Robust Stabilization in the Gap-topology (Lecture Notes in Control and Information Sciences) issue 150

Lecture Notes in Control and Information Sciences Edited by M.Thoma and A.Wyner

150 L. C. G.J.M. Habets

Robust Stabilization in the Gap-topology

Springer-Vedag Berlin Heidelberg New York London Paris Tokyo Hong Kong Barcelona

Series Editors M. Thoma • A. Wyner

Advisory Board L. D. Davisson • A. G. J. MacFarlane - H. Kwakernaak J. L. Massey • Ya Z. Tsypkin • A. J. Viterbi

Author Luc C.G.J.M. Habets Eindhoven University of Technology Dept. of Mathematics and Computing Science Dommelbuilding 1.11 RO. Box 513 5600 MB Eindhoven The Netherlands

ISBN 3-540-53466-0 Spdnger-Vedag Berlin Heidelberg NewYork ISBN 0-387-53466-0 Spdnger-Vedag NewYork Berlin Heidelberg This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in other ways, and storage in data banks. Duplication of this publication or parts thereof is only permitted under the provisions of the German Copyright Law of September 9, 1965, in its current version, and a copyright fee must always be paid. Violations fall under the prosecution act of the German Copyright Law. © Spdnger-Vedag Bedin, Heidelberg 1991 Printed in Germany The use of registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Printing: Mercedes-Druck, Berlin Binding: B. Helm, Berlin 61/3020-543210 Printed on acid-free paper.

Crescit scribendo scribendi studium (Desiderius Erasmus)

PREFACE

'Crescit scribendo scribendi studium'. This experience I had, while writing this book. When I started, I thought I had everything clearly in mind, but, while writing, things turned out to be more complicated. The obligation to state results precisely, forced me to treat the subject in more detail than I planned. I hope however that a greater insight of the reader in the various notions and subdlity of the theory will be the gain of this effort. But it took me more paper to tell the story than I expected, and I hope that this f'mal result is clarifying instead of tedious or boring to the reader. Roughly speaking, this book can be divided into three main parts. In the ftrst one, the notions of robust stabilization and the gap--topology are introduced. In the second part, most of the

theory is developed. We

find

sufficient conditions for robust

stabilization

in

the

gap--topology, and are able to derive a method for the solution of the problem of optimally robust control. Finally, in the third part, algorithms are developed to actually compute a solution to this problem for systems in state-space form. So in this book we slowly tend from theory to practice. Of course I'm grateful to all the people who helped me, in one way or the other, with the accomplishment of this paper. First of alt I'd like to thank prof. M.L.J. Hautus. He awoke my interest for the field of systems theory and also attended me to this subject. I am also indebted to Siquan Zhu, who developed most of the theory, upon which my algorithms are based. Especially I would like to thank dr. Kees Praagman

for carefully reading the

manuscript and all his valuable suggestions. The cooperation with him was very stimulating and I'd like to thank him for always being available when I had questions or problems. I'm also grateful to drs. Anton Stoorvogel. When I didn't understand the computer anymore, he always got me out of trouble. Finally I'd like to mention drs. J. Timmermans. Without him, this book would probably have never been written. And last but not least, I thank my parents for all their support and encouragements. Without them I'd never come so far. Luc Habets.

CONTENTS

0.

Introduction

1.

Robust stabilization

3

1.1.

Stability and stabilizability

3

1.2.

Bezout factorizations; parametrization of all stabilizing controllers

1.3. 2.

5

Robust stabilization

10

The gap--topology

12

2.1. 2.2.

The gap between two closed subspaces The gap-topology

13

2.3.

A necessary and sufficient

condition for robust

stabilization

3.

4.

15

2.4.

The gap--metric

18

2.5.

An other topology for plants

20

Sufficient conditions for robust stabilization

21

3.1.

The gap-metric and Bezout factorizations

21

3.2.

Guaranteed bounds for robust stabilization

25

Optimally robust control

29

4.1.

Normalized Bezout factorizations

29

4.2.

Optimally robust control

41

4.3.

Computation of an optimally robust controller and its stability radius

5.

12

45

An algorithm to compute an optimally robust controller and its stability radius

51

5.1.

The non-dynamical case

51

5.2.

A state-space realization of the matrix V

52

5.3.

Computation of the optimal stability radius wit; Hankel---operators

59

5.4,

Solving the Nehari-problem

65

5.5.

An algorithm to compute an optimally robust controller and its stability radius

71

VI 6.

7.

Reduction of the order of the compensator

75

6.1.

The order of the compensator

75

6.2.

The method of Glover

77

6.3.

The method of Glover and McFarlane

81

6.4.

The performance of the algorithms

86

Conclusions

91

7.1.

Summary

91

7.2.

An outline for future developments

92

Appendix A

97

Appendix B

102 107

Appendix C Appendix D Appendix E

111

Appendix F

120

References

125

116

VII NOTATION

Given a state-space realization [A,B,C,D] of a plant P, the transfermatrix corresponding to P is given by P(s) = D + C(sI-A)-IB. Throughout the text we use the notation [A,B,C,D] not only to describe the realization of P, but also to indicate the related transfermatrix, i.e. [A,B,C,D] = D + C(sI-A)-IB. Using this

data---structure we

can easily obtain

the following formulae

transfermatlices.

[A,B,C,D] = [T-IAT,T-IB,CT,D]. [A,B,C,D]-I =- [A-BD-IC,BD--t,-D--IC,D--I]. [AI,BI,CI,D I] + [A2,B2,C2,D 2] = --

B2

[A1,BI,CI,D11 * [A2,B2,C~,D2] =

for operations

on

O. I N T R O D U C T I O N

This book addresses the problem of robust stabilization. Given a plant P

we often

encounter the problem of stabilization: find a compensator C such that the transferfunction of the closed-loop-system is stable. But when P is stabilizable, there is often a whole set of stabilizing compensators.

Now we

get the problem which one to choose. So we need a

criterion to select a certain stabilizing compensator C. In practice however, the description of the plant P is often inaccurate (for example because the plant--description is approximated by statistical means). So we don't work with the real plant P, but with a model of it, say Pro- Now, when we construct a compensator C, that stabilizes Pro, the question is whether it stabilizes P

also. When

P

and Pm differ only

slightly, we expect C to do so. But how much P and Pra may differ? Now we can describe the problem of robust stabilization. Given a plant P, we want to find a compensator C, that not only stabilizes P, but also a neighborhood of P. By changing C,

we

will

also

change

the

stabilized

neighborhood.

Now

we

are

interested

in

that

compensator C, that stabilizes the largest neighborhood of P. In practice this means the following. Suppose we have a (probably inaccurate) model Pm of the plant P. We stabilize Pm by that compensator C, that also stabilizes the greatest neighborhood of Pro. So, when the real plant P is in that neighborhood, it is also stabilized by C. In this way we have designed a robust compensator C. Now we will give a short outline of the contents of this paper. In the first chapter we will

make

the

above

statements

more

precise.

We

give

a

definition

of

stability

and

parametrize all stabilizing compensators. At the end we give a more formal description of the problem of robust stabilization. In the second chapter we introduce the gap-metric. This metric gives a measure for the distance

between

two plants.

The

gap-metric,

and

the gap-topology,

induced

by this

metric, will turn out to be very important for the robust stabilization problem, because they determine whether two plants are close to each other or not. With the help of the gap-metric, we are able to derive sufficient conditions for robust stabilization. This is the content of chapter 3. Using

the

bounds

for robust

stabilization

found

in chapter

3, we

will tackle

the

problem of optimally robust control in chapter 4. We derive a method that gives an optimally robust compensator for each plant P. In chapter 5 we translate the theoretical method of chapter 4 to an algorithm, that actually computes a solution of the problem, for systems in state-space form. Unfortunately the algorithm, described in chapter 5, gives solutions

of a very high

order. This means that in practice the solution is hardly applicable. In chapter 6 we describe

an algorithm that gives solutions of a much lower order. Also an other method will be discussed, which decreases the order of the solution drastically. Finally in chapter 7 we draw our conclusions and give a short outline for eventual future developments in this field. At the end of this intxoduetion we z~mark that, certainly in the first part of this text, we follow along the same lines as Zhu in [19]. In this dissertation however, Zhu treats the same problem in a more general context. We eonf'me ourselves to the ease that all the entries of the transfermatrix of a plant are rational functions. We often even assume these functions to be real--rational (this ease represents the eontinuous--~me lumped LTI-systems).

I. ROBUST STAB-.rZATION In this chapter we introduce the

concept

factorizations we

are

of

first give definitions of Bezout

able

to

factorizations of

give conditions for

stability and a

plant P.

the

stabilizability. Then With

the

stabilizabity of

a

help

we

of these

plant

P,

and

parametrize all stabilizing compensators. Finally we give a more precise characterization of the problem of robust stabilization. 1.1. Stability and stal~liT~bility Let 8(s) denote the set of all rational functions in s with real coefficients, and M(~(s)) the set of all matrices with elements in 8(s). So M(R(s)) consists of all matrices with real-rational

functions

as

entries.

Throughout

this

text

we

usually

assume

that

the

transferfunction P(s) of a plant P is an element of M(~(s)). By ~p(S) we denote the set of all proper rational functions in s with real coefficients. So 8p(S) is a subset of 8(s) consisting of all rational functions in 8(s), for which the degree of the numerator is not larger than the degree of the denominator. Again MORp(S)) consists of all matrices with elements in Rp(S). In the last chapters of this paper we will assume that the transfermatrix P(s) of a plant P belongs to M(~p(s)). Now we introduce the space H**, which plays an important role in the problem of robust stabilization. DEF.I.I.1. The space H

consists of all functions F(s) of the complex variable s, which are

analytic in Re s > 0, take values in f, and are bounded in Re s > 0, i.e. sup [ I F(s) ] ] Re s > 0 } < ** The left--hand side of (1.1) defines the /-/ - n o r m of a function F ¢ H .

(1.1) With this norm, H,,

is a Banach-space. It is a well known fact that we can replace the open right half-plane in (1.1) by the imaginary axis. This gives the following, mostly used formula for the norm of a function F H**:

II W II~ = sup { I F(io~) [ [ co e • }

(1.2)

DEF.I.I.2. ~-,I-I** := ~(s) t~ Note that F~H** is a subspace of H**. We also remark that ~,I-I, is a subspace of ~p(s), because a real rational function in H** has to be proper. Let, analogous as before, M(H**) and

M(RH**) denote the sets of all matrices with elements in H** and RH** respectively. Then we can now give a definition of stability: DEF.I.I.3 A plant P is stable when its transferfunction P(s) is an element of M(H). Under the assumption P(s) e M(~(s)), (1.1.3) becomes: a plant P with transferfunction P(s) e M(~(s)) is stable if P(s) e M(RH**). Remark that the notion of stability we have introduced here, is BIBO--stability (bounded-input, bounded--output). A real-rational transferfunction is stable if it is proper, and all its poles have negative real parts. We now introduce the feedback-system shown in figure (1.3), where'P represents the plant (with transferfunction P), and C a compensator (analogous to the plant P, we assume that the entries of the transfermatrix C of the compensator C are all real--rational functions, so C e M(~(s)) ).

u2[+

Y2 (fig.l.3)

upu2 denote external inputs, el,e 2 inputs to the compensator and the system respectively, and Yt and Y2 outputs of the compensator and the plant respectively. This model is versatile

enough to accommodate several control problems. Later on we will refer to this set--up as a feedback-system. el ] is given by Suppose P,C e M(~(s)). The transfermatrix from u := [ ul uz ] to e := [ ez (see [15, p.101]): H(P,C) := [ ( I + P C ) - t - P ( I + C P ) - I ] C(I +PC) -1

(1.4)

(I+CP)-t

where we assumed that P and C have compatible dimensions, and the weU--posedness condition [I+CP[ = [I+PCI e 0, is also satisfied; so H(P,C) makes sense. In this case, the transferfunction W ( P , C ) f r o m u t o y :=/ YI ] i s L Y2 J

given by:

(1.5) We can now give the following definition of stabilizability:

DEF.I.I.4. A system P e M(~(s)) is said to be stabilizable if there exists a compensator C e M(~(s)) such that

II+PCl

,~ 0, and the transferfunction W(P,C) from u to y is stable. If

W(P,C) is stable, C is called a stabilizing compensator of P. From formula (1.5) we see immediately that W(P,C) is stable if and only if H(P,C) is stable. Because H(P,C)

has a

simpler form than W(P,C),

we

mostly use the following

characterization of stabilizability (the proof is obvious from the argument above). THEOREM 1.1.1. A system P e M(~(s)) is stabilizable iff there exists a compensator C e M(~(s)) such that II+PCI ~ 0 and the transfeffunction H(P,C) from u to e is stable. Finally we remark that the conditions for stability are symmetric in P and C, i.e. H(P,C) is stable iff H(C,P) is stable. This because

[ 0i - 0I ] H f P , C ) [ _ 0

I ] = H(C,P)

(1.6)

1.2. Bezout factorizafions; parametrization of all stabilizing controllers In this section we introduce the concept of right- and left-Bezout factorizations. With help of these factorizations we can give conditions for the stabilizability of a plant P. Finally we are able to parametrize all compensators C, that stabilize a certain plant P. DEF.1.2.1. Let

P an nxm transfermatrix, P

•

M(N(s)).

We

say (D,N)

e

M(~H,)

is a

right-Bezout factorization (r.b.f.) of P if: 1) D is an mxm, and N an nxm transfermatrix, and [ D t = det(D) ~ 0; 2) There exist two matrices Y and Z in M(~-I.) such that YD + ZN = I 3) P = ND-1

(1.7)

Similarly we can define left-Bezout factorizations:

DEF.1.2.2. Let

P an nxm transfermatrix, P

•

M(~(s)). We

say (lq,13) e

M(SH)

left-Bezout factorization (1.b.f.) of P if: 1) ~ is an nxn, and ~ an nxm transfermatrix, and [ ~

[ = det('lS) ;~ 0;

2) There exist two matrices ~ and 7. in M(~r-I,.) such that

59

+ ~2 = I

3) P = f3-1~

(1.8)

is a

Remark that an r.b.L (resp. 1.b.L) of a P e M(~(s)) is unique up to right (resp. left) multiplication by unimodular matrices. In our case (P • MOR(s))) the concept of Bezout factofizations is completely analogous to the notion of coprime factorizations introduced by Vidyasagar in [15, pp.74-75].

Since in

our case there is no difference between these two ideas, we can use both these terms to point out the same concept. We can now ask the question: when does a transfermatrix P e MOl(s)) have a Bezout factorization? The following theorem gives the answer. THEOREM 1.2.1. Each P e M(R(s)) has a right- and a left-Bezout factorizafion. PROOF The proof of this theorem can be found in [15, p.75]. A constructive proof for the case P

e

M(~p(s))

is given in

[4, pp.23-25].

Later on, in chapter 4, we will give a

constructive proof of the existence of so called normalized Bezout factorizations for this case. These

are

a

special

sort Bezout

factorizations.

So for P

e

M(~p(s))

that proof is also

sufficient for the existence of ordinary fight-- and left-Bezout factorizations. For the sake of completeness we mention the fact that for other (more general) classes of transfermatrices, the existence of f i g h t - or left-Bezout fractions is not always guaranteed (for example f(s) = se--s

doesn't have a fight- or left-Bezout factorization in H**).

With the help of the concept of Bezout factofizations we can now give a condition for the stabilizability of a plant. THEOREM 1.2.2. If the transfermatrix P of a plant P has a right-Bezout factofization, then P is stabilizable. Moreover, any stabilizing controller C of P, has a left-Bezout factorization. PROOF First we show that if P has an r.b.f, then P is stabilizable. Assume (N,D) is an r.b.f. of P. Now suppose there are matrices Y and Z such that YD+ZN=U is invertible. Suppose [ Y I ~ 0. When we now define C := y-1Z, then

(1.9) We prove (1.9) by checking the four equalities of formula (1.4). 1)

(I+PC) -q = I - N(YD+ZN)qZ.

We show (I+PC)(I-NU-IZ) = I. (I+PC)(I-NU-IZ) = (I+NI~IYqZ)(I-NU-IZ) =

= I-NU-aZ+ND-ty-aZ-ND-Iy-IZNU-aZ = I -

N(U-LIY-IY-I+D-ty-IZNU-t)Z =

= I-N(U-L--D-ty-I+D-Iy-I(U-YD)U-I)Z = I -

N(U-a--D-IY-t+D-aY-u-U-t)Z = I.

2)

C(I+PC) -I -- D(YD+ZN)-IZ.

C(I+PC)-I -- y - I Z ( I - N U - I Z ) = Y-IZ = y-iz 3)

Y-IZNU-IZ = y - t z

-

y-I(U-YD)U-IZ =

_ y - 1 Z + DU--aZ = D(YD+ZN)-aZ. (I+CP) --I = DCYD+ZN)-IY.

W e show (I+CP)DU-IY = I. (I+CP)DU-IY = (I+Y-IZND-t)DU-IY = D U - I y + Y - t Z N U - I Y = = DU-IY + Y - I ( U - Y D ) U - I y = D U - t Y + I - D U - I Y = I. 4)

-P(I+Cp)-t = -N(YD+ZN)-t.

-P(I+CP)-t = -ND-tD(YD+ZN)-IY = -N(YD+ZN)-tY. Now

we

know

that

(N,D)

is an r.b.f, o f P.

So there exist matrices Y

and

Z

in

M(P,I-/**) such that YD + Z N = I. When

I Y

I ** 0, w e see immediately that w h e n w e c h o o s e C = Y-aZ, then

Because N,D,Y and Z are all matrices in M(PJ-I**), H(P,C) e M(P,I-I ) and so C stabilizes P. N o w suppose

[ Y

[ = 0. Let K be a matrix in M(P~-I ) o f the same size as Y and

such that ( y T , K T ) T has full column rank. Define

V :-- { R e M(P,H**) [

I Y+RK

I *0

}.

Then V is an open and dense subset o f the set o f all malaSees in M(P,I-I.,) o f the same size as Y (for a proof, see [15, p. 111]). N o w w e take an R e V Then (I + R K D )

is invertible, with bounded

such that

I! RIO) 11

inverse (for a proof, see l e m m a

3.2.2.)

< 1. and

because o f the B e z o u t - i d e n t i t y (1.7) we have (Y+RK)D + ZN = RKD + YD + ZN = I + RKD.

Since

(I +

RKD)

C := (Y+RK)-IZ,

is invertible and

[ Y

+

RK

[ ~

0, we

have,

when

we

choose

Because all the matrices on the tight-hand side belong to MOLtt**), H(P,C) e MORH,~), and C stabilizes P. The second part of the proof, the fact that each stabilizing controller has an 1.b.f., is quite clear from the exposition above. For an exact proof, however, we refer to [15, p.363]. o Analogous to theorem 1.2.2., the following holds: if P has a left-Bezout factorization, then P is stabilizable, and any stabilizing compensator C has a right-Bezout factorization. From theorems 1.2.1. and 1.2.2., we obtain immediately: COROLLARY 1.2.3. Each plant P ~ M(~(s)) is stabilizable, and any stabilizing controller C has an r.b.f, and an l.b.f. Now we have seen that each plant P ~ M(~(s)) is stabilizable, we want to know how this can be done. To this end, we will now parametrize all stabilizing compensators of a certain plant P. Let S(P) denote the set of all compensators C that stabilize P. Assume that (D,N) and (~,13) are an r.b.f, and an 1.b.f. of P respectively. Let C O e S(P), and (Z,Y) and (~,~) be an 1.b.f. and an r.b.f, of C o respectively, such that:

I -z YI I -N ]D

N

D

2

=

EI °l 0

I

(I,10)

From (1.10) we obtain for any rnxn transfermawix R e M(~H**):

b

Y-R 1 [-N N

D

2+DR

=

[i o 1 0

(1.11)

I

With formula (1.11) it is not very difficult to prove the following parametrization

theorem

(for an exact proof, see [15, p.108]).

THEOREM 1.2.4. Suppose P e M(~(s)), and let (D,N) and (1~,13) be any r.b.f, and 1.b.f. of P. Select man-ices Y,Z,~,~. ~ M(~ffI.) such that YD + ZN = I, D ~ + ]q~. = I. Then:

S(P) = { (Y-Rlq)-1(Z+R]3)

I R ~ M(RH**), { Y - R ~ I , 0 } (1.12)

= { (2+DR)(~-NR)-~

I R ~ M(RHo), I ~ - N R I

~ 0 ]

9

Finally we prove the following result:

THEOREM 1.2.5. Suppose P ~ Mt~(s)) and let (D,b0 and flq,13) be any r.b.L and 1.b.L of P respectively. Let C o ~ S(P) and (Z,Y) and (~,~) an 1.b.f. and an r.b.f, of C O respectively, such that (1.10) holds. Then we have for any nxm l~ansfermatrix R G M(IRH**):

1) I Y-Rlq I ~ 0 if and only if I ? - N ~ I

* 0;

2) If I Y-R2q I * 0, then (Y-R_~)~(Z+R~) = (~+DR)(~-NR)-I

PROOF 1) Assume

I Y -

(1.13)

R/q ] ~ 0. According to theorem 1.2.5., the compensator C =

(Y-Rlq)-I(Z+R_I3) is an element of S(P). Let (Ye, Ze) be an r.b.f, of C. Again using theorem 1.2.4, we see that there exists an R e • M(RH**) such that } ~ - NR e [ ~ 0 and (Y¢,Zc) = (~-NRc,Z+DRc). So

( Y - P ~ ) - I ( Z + R ~ ) = ZeYEl = (~,+DRe)(?-NRc)-I.

(1.14)

Pre--multiplying (1.14) with (Y-ILK/) and post-multiplying it by ( ~ - N R e) gives

(Z+R[3)(~-NRe) = ( Y - I ~ ) ( ~ + D R e ) , so Z Y - Z N R e + R . ~ - R I 3 N R c = Y 2 + Y D R c - R ~ Z - P , aqDR e. Rearranging gives:

R(~5~+lqZ) - (ZN+YD)R c = ( Y 2 - Z ~ ) - R(IqD-I3N)R e.

But by formula (1.10) we know that (15~+1q2) = I, (ZN+YD) = I, YT.-Z~ = 0 and 5~D-bN = 0,

so R - R e = 0, and R = R c. Hence [ ~ - NR [ ~ 0. The converse can be proved completely analogously. 2) The

proof of the

second

statement

follows already

from

the

Combination of (1.14) and the fact that R = R c immediately gives (1.13). []

arguments

above.

10 1.3. Robust stabilization

In this section we formulate the problem of robust stabilization, described in the introduction, more precisely. Suppose we have a sequence {Pz} of systems and a sequence {C.L] of controllers, parametrized by ~ taking values in a metric space A. Suppose H(P0,C 0) is stable. The central question is: when will H(Pk,C~,) also be stable as X is sufficiently close

to O, and H(Px, C~,) , Hff'0,C0) as ~ , 0. Here the space A of parameters can arise from perturbations, disturbances, modelling errors, measurement errors, etc. In practice this means the following. Po is the nominal system or a mathematical model (Pro in the introduction), which approximately describes the unknown real physical system. C O is the nominal (ideal) controller, designed according to the nominal plant P0. The nominal system P0 and compensator C o form a stable pair, i.e. H(P0,C0) is stable. Now we hope that in practice the real physical system PZ' which is close to Po, and the real controller CX, which is close to C o, also form a stable pair, and that H(P~,C~.) is close to H(P0,C0). This stabilization.

is the problem of robustness of feedback stabilization,

or

simply robust

We even want to go a step further. In section 1.2. we saw that there is not one stabilizing controller for a plant P, but a whole set: S(P). Now we are interested in that controller C that meets the robustness--criterion the best. In other words, we want to find the optimally robust controller C. That is the controller from the set S(P), that not only stabilizes P, but also the largest neighborhood of P. To make the last statements more precise we have to introduce a topology or a metric for plants, in order to describe the distance from P0 to PZ,' and from C o to C~.. The operator--norm is not applicable for this purpose, because this norm is not defined for unstable systems.

However,

stabilization, in the

we

require

that

this

topology

sense that perturbation

is

(Pz,CZ)

compatible from

with

(P0,Co)

is

the a

robustness of

stable pair,

and

H(P2L,C~) is close to H(P0,C0) if P~. is close to P0 and C~, is close to C o in the topology. The foUowing topology meets this requirement. Let Rn*=(s) denote the set of all nxm matrices with elements in ~(s). First we define a basic neighborhood N of P0 ~ ~n*m(s):

N := N(Po,Co,e) := [ P e gqn*m(s) I H ( P , C o ) is stable and II H(P,Co)--H(Po,Co) II < e }

(1.15)

where C o is one of the stabilizing controllers of P0" By varying e over ~+, C O over S(P0), and P0 over ~n*m(s), we obtain a collection of basic neighborhoods, which forms a basis for a topology T over Rn*m(s).

Although

this

topology perfectly describes

the robusmess of

feedback

stabilization,

unfortunately it has too little structure to be studied from this definition. That is why, in the

11 next chapter, w e

will introduce another topology, the gap---topology, which in the end will

appear to bc identical to the topology T on ~n*m($).

2. THE G A P - T O P O L O G Y In this section we introduce a topology for plants: the gap-topology. First we define the gap between two closed subspaces of a Banach--space X. With the aid of this concept, we can describe the gap---topology for plants. Then we first quote some results about the gap-topology from other authors, before we derive a necessary and sufficient condition for robust stabilization. Finally we show how the gap-topology can be metrized, and how this so called

gap---metric

graph-topology,

can

be

computed.

In

another topology for plants,

the

at

the

suggested by Vidyasagar in [15, sec.7.2],

last

section

we

shortly

glance

and

compare this to our gap-topology. 2.1. The gap between two closed subspaces Let X be a Banach--space and q~ and V two closed subspaces in X. We define

sq,:=

{ x~

~ I I I x [I = 1 )

Now the directed gap from ~ to ~ is given by

8(q~,R/) := sup xeScp

inf ye~

(2.1)

H x - y II

In order to measure the "distance" between two closed subspaces, the gap between q~ and is now defined as:

8(cp,~) := max { g(tp,~) , 8(~,~) }

(2.2)

Moreover, we define

~(0,~p) := 0. From these definitions and the fact that ~(cp,0) = 1 if and only if {p # 0, we get directly the following relations: 8(cp,~g) = 0 iff ~p = ~;

8(q~,~/) = 8W,cp);

0 < 8(q~,~) < 1 Remark

that the gap $(.,.) is in general not a metric on the space of all dosed

subspaces of X, because the triangle inequality is not always satisfied.

13 2.2. The gap--topology Before we introduce the gap-topology, we first look at the transfermatrix P of a plant P in a slightly different way. Suppose P ¢ ~a*m(s). Then P transforms an m-dimensional input to an n-dimensional output. In our case the i n p u t - and output-space are, apart from dimension, the same: the Banach-space H 2.

DEF.2.2.1. The Hardy-space

1~2 is the space of all functions x(s) which are analytic in

Re s > 0, take values in (n, and satisfy the uniform square-integrability condition:

II x I1= := [ ~>0 sup

(2~)-t ~ l~ II x(g+ie0) II2 dco ] t/2 < **

(2.3)

(2.3) defines a norm on H~, with this norm H~ is a Banach-space.

(Because H~ is a closed subspace of the I-filbert-space L~, H~ is a Hilbert-space itself.) Each transfermatrix P(.) e ~n*In(s) now induces an operator P: the domain Dom(P) of P consists of all elements x(.) in H~ such that the product P(.)x(.) is in H~, the action of P on x(.) e Dora(P) is defined as P(.)x(.). In this way we see a transfermatrix P(.) ~ ~'*m(s) as an operator P, mapping a subspace of H~ into H~. For the operator P, we can define the graph of P as: G(P) := { (x,Px) [ x e Dom(P) }.

(2.4)

In [18, p.850] Zhu proved that for each plant--compensator pair (P,C) e M(~(s)), such that H(P,C) is stable, the induced operators P and C are closed. Since we know (corollary 1.2.3.) that each P ~ M(8(s)) is stabilizable, it follows immediately that for each P ~ ~n*m(s), the induced operator P is closed. So G(P) is a closed subspace of (Xm x Xn). Now we can introduce the gap between two systems PI and P2. Suppose P1,P2 e Rn*m(s). The gap between P1 and P2 is defined as the gap between their graphs respectively, i.e. ~(P1,P2) := ~(G(PI),G(P2)). From this formula follows immediately that 8(P1,P 2) = 0 if and only if P1 -- P2"

A basic neighborhood of P0 ~ ~"*m(s) we define as:

(2.5)

14 N(P0,e) := { P e ~n'm(s) I 8(P'P0) < ~ }"

By varying e over (0,1] and varying P0 over ~n*m(s), we obtain a collection of basic neighborhoods. This collection is a basis for a topology on ~"*m(s), which we call the

gap-topology. We now give some properties of the gap-topology, quoted from other authors. The lemma's 2.2.1.-2.2.4. come from Kato ([9, pp.197-206]), lemma 2.2.5. from Zhu ([18, p. 8521).

LEMMA 2.2.1. If P0 is an nxm matrix, P0 • MtRH ), and P • ~n*m(s) satisfies ~(P,P0) < (1 + ]1 P0 H2)-112 then P a M(RI-I ).

So

the

set of

nxrn

matrices

in

M(r,I-I)

is

an

open

subset

of ~{n*m(s) in the

gap-topology. As a consequence, any system is stable when it is sufficiently close to a stable

one. THEOREM 2.2.2. On M(~I-I ) is the gap-topology identical to the topology induced by the operator--norm.

LEMMA 2.2.3. Let P],P2 G l~n*m(s), and Po an nxm matrix in M(IRI-I,). Then: ~(PI+P0,P2+P0) < 2 (I + II P0 II2) 8(Pl,P2)" LEMMA 2.2.4. If P1,P2

• }{n*n(s) are

(2.6)

invertible, then

80~11,~2 I) = 8(P1,P2).

(2.7)

LEMMA 2.2.5. Suppose that Pi • l~n*m(s) (i=1,2) have the following diagonal form:

Pi = IP~ 0 p2O] Then:

(i=1,2).

(2.8)

15

max { B(PI,P~),BCP],P ~) } < B(P122) < B(Pl,P~) + B(P2,P~). From lemma 2.2.5. we immediately derive (the proof is obvious):

COROLLARY 2.2.6. Let {P~} be a sequence of systems in ~*n(s), with the diagonal form:

o

P~ , 0) if and only if 5(P~,PoI)

Then, ~(P~.,Po) , 0 (as ~. simultaneously (as ~ , 0).

, 0 and BCP~,P02)

, 0

We will use this so called diagonal product property in the next section.

2.3. A necessary and sufficient condition for robust stabilization In this section we apply the gap--topology to the problem of robust stabilization. We show that the gap--topology is compatible with this problem, and is identical to the topology T, given in section 1.3.

THEOREM

2.3.1. Let

{P~}

c ~n*m(s) be a sequence of systems, and

{C}} c ~m*"(s) a

sequence of controllers. Then: ~i(H(P~,C.L),H(Po,Co)) ~

0

(~.

, 0)

iff 8(P~.P o) ----* 0

(~.

, 0);

~(dk,C o)

, 0

(~

simultaneously. PROOF It is not difficult to show that H2t := H(Px,C~L) can be written as: H~. = (I + FO~L)q with

F :=

ioii -I 0

O~, :=

0

o1

PX

, O)

16 (see for example [15, pp.100-101]) According to lemma 2.2.4. we have

8(H~,Ho) = 8( fI+FG~)-: , (I +FGo)-I ) = 5( fI+FG~), (I+FG o) ).

(2.9)

Now the application of (2.6) (two times) gives: 1/4 8(FG~,FG o) = = l/4 5((I+FGk)+(-I) , (I+FGo)+(-I)) -< 5((I+FG~) , (I+FG0)) -< 4 15(FG~,FG0).

(2.10)

(2.9) and (2.10) together yield :/4 8(FGx,FGo) < 8(H~,Ho) _< 4 8(FGk,FGo).

(2.1I)

Analogous to 2.2.5. it can be proved that lemma 2.2.5. still holds when Pi 0 =1,2) is defined as

o

Vi :=

P~

0

(i=1,2).

This implies that max {8(P?dPo),5(C~,Co)}

_< 8(FG~,FG0) _< ~5(Pk,Po) + 8(Ck,Co).

Combination of this last formula with (2.11) yields 1/4 max {8(P~,Po),5(C~.,Co)}

< 8(H)~,Ho) < 4 [8(P~,P 0) + 8(Ck,Co)] (2.12)

From (2.12) our claim follows immediately, rn With help of theorem 2.3.1. we can now give a necessary and sufficient condition for robustness of feedback stabilization. COROLLARY 2.3.2. Consider the feedback system in fig.(l.3). Suppose Po and C o are the nominal system and controller respectively, and H(Po,Co) is stable. Let {P~.]X>0 and {C).}L>0 be families of systems and controllers respectively. Then:

17 3 ~o > 0: V ~. < ~-0: H(Pk,C~) is stable, and

II H(P%,C k) -n(Po,C0) II

, 0

(~.

, 0)

(2.13)

if and only if 8(P)-Po) - - - , o

(x

, o)

~ ( c ~ , c o)

, o

(%

, o)

simultaneously.

PROOF (~.

"#" Suppose H(P~.,Ck) is stable for k < Xo, and II H(P~,C%) -H(Po,Co) II ,o , 0). H(P~,C~) is stable. According to theorem 2.2.2., the gap-topology is identical to

the topology induced by the operator norm. Therefore we have: ~(H(P~,C~),H(P0.C0)) (~. , 0). Now application of theorem 2.3.1. gives the necessity.

, 0

"¢" Suppose ~(P~,P0) , 0 (~. , 0) and 8(C~,C 0) , 0 (~. , 0) simultaneously. Then by theorem 2.3.1. we have 8(H(P%,C;L),H(Po,C0) ) , 0 (~. ~ 0). According to lemma 2.2.1. H(Px,C~) is stable as % is sufficiently close to 0 (because then H(P~.,C~) is close to H(Po,Co), which is stable). Again application of theorem 2.2.2. gives that 5(I--I(P~,C~),H(P0,C0) ) , 0 (~. ; 0) now implies I[ H ( P ~ , C ~ . ) - H(P0,C0)II , 0 O. ,0). n

Finally we

show that the gap-topology on ~n*m(s) is identical to the topology T,

defined in section 1.3. In this toplogy, a basic neighborhood of P ~ ~n*m(s) was defined as

N(P0,Co,e) := { P • ~n*m(s) [ H(P,C 0) is s t a b l e

and

II H ( P , C o ) - H(Po,Co) II < e } with C O one of the stabilizing compensators of Po. Now suppose the sequence {P~.} c ~n*m(s) converges to P o e Rn*m(s) in the topology T. Then we have that H(P~,Co) is stable when ~. is sufficiently close to 0, and II H(Px,C0) H(P0,C0) II , 0 gap-topology also.

(~

, 0). So according to corollary 2.3.2., {P~} converges to Po in the

Conversely, suppose {P~} ¢ Rn*m(s) converges to Po ~ Rn*m(s) in the gap-topology. Suppose H(Po,Co) is stable. Then, by corollary 2.3.2., we know that H(Pk,Co) is stable when % is sufficiently close to 0, and II H(P~.,C0) - H(Po,Co) II , o (~. , o). This means that {P%} converges to Po in the topology T.

18 2.4. The gap-metric In section 2.1. we saw that the gap 8 between two closed subspaces is in general not a metric. In this section we show that in our case, where the space H 2 of inputs and outputs is a Hilbert--,space, the gap 5 between two plants is in fact a metric. We also give a method to compute this gap--metric in practice. Suppose P1,P2 ~ Rn*m(s), and we regard P1 and P2 as operators from a subspace of the Hilbert--space H~ to the Hilbert-spaee H~. Let TI(P.~ (i=1,2) denote the orthogonal projection from ~

x H E onto the graph G(P-~ of Pi (i=1,2). Then it is not difficult to see that

8(PvP2)

= sup XeSG(p1 ) =

sup

inf [Ix - y [1 = YeG(P2)

II (I-ncP2))x II

=

XeSG(p1) = sup [I (I-FI(P2))FI(Pi)x X¢SG(Pi)

II

=

-- sup II (i-n(P:))II(P~)x II n xaH2m ~ H2, Ilxll=1 = II (l-n(v2))rl(P,)II-

=

(2.14)

With this formula it is shown (see [9, pp. 56--58], or [10, p.205] for a direct proof) that ~(Pa,P=)

=

II

17(P,) -II(P2) II-

(2.15)

From formula (2.15) it is clear that 8(.,.) defines a metric for plants. We call this metric the

gap--metric. To calculate the gap--metric, we need to f'md a representation of YI(P) for P ~ M(~(s)). Such a representation can be given with help of Bezont factorizations. However, we need the following preliminary lemma (for a proof, see [20, p.55]).

LEMMA 2.4.1. Let P e lln*m(s), and suppose (D,N) is an r.b.f, of P. Let D* and N* denote the

adjoint

operators

of

N* = NT(---s) ). Then S := D * D + N ' N

has a bounded inverse.

D

and

N

(in

l.~---sense)

respectively

(so

D* = DT(---s),

19 Analogously, if (~j3) is an 1.b.f. of P, then S := (1313' + lqll~*) has a bounded inverse. With this knowledge the following theorem is not difficult to prove. In fact, it is a direct consequence of lernma 3.14 in [3, p.704]. THEOREM 2.4.2. Let P ~ ~n*m(s), and (D,N) and (lqJ3) an r.b.f, respectively. Then:

and an 1.b.f. of P

YI(P) = [ D ] (D*D +N*N-)-I (D*,N*)

= I --

-13" ] (~i13" +~l~*)-t

lq*

(-15,1q).

(2.16)

J

With help of the formula's (2.15) and (2.16) it is in principle possible to calculate the gap between two plants. Another, probably better, method to compute the gap---metric was introduced by Georgiou (see [5, p.254]). In this article he proved the following theorem. THEOREM 2.4.3. Let PI,P2 ~ ~n*m(s), and (D1,NI) and (D2,N2) nomuzlized right-Bezout factorizations of PI and P2 respectively. Then:

~(PI'P2) = Q einfM(I-I**) [ I [ DIN1] - [

D2N2]Q [[.

(2.17)

An exact definition of normalized right-Bezout factorizations will be given in chapter 4. We here only mention that it is an r.b.f, with the extra property D*D + N*N = I and that such a factorization for a P ~ ~n*m(s) always exists (see [15, p.262]). From theorem 2.4.3. we see that the computation of the gap between two plants can be reduced to the computation of two infima as given in formula (2.17). For P e M(~p(S)) such infmaa can be computed with the method of Francis, described in [4, oh.8]. In a later article [6, pp.5-6] Georgiou applied this method to this problem, and worked it out. So this later article gives a (theoretical) algorithm to compute the gap-metric. For computer purposes this algorithm is implemented in the MATLAB-function gap. For the details of this implementation we refer to appendix A.

20 2.5. An other topology for plants In this last section of chapter 2, we want, for the sake of completeness, mention an other topology for plants: the graph-topology. This topology was introduced by Vidyasagar in [15, sec.7.2.]. For a definition and properties, we therefore refer to this book. Zhu however proved (see [18, pp. 853---854]) that on ~a*rn(s) the gap- and graph-topology are identical. Although

the

gap-

and

the

graph--topology

are,

in

our

case,

the

same,

the

gap---topology has one great advantage: the gap--mewic is much easier to compute than the graph-metric. In the last section we mentioned the method of Georgiou, that really solves the computation problem for the gap-metric. The graph---metric, on the other hand, is much more difficult to compute. Another advantage of the gap-topology, which is not so important to us, is that when one works with more general systems, the gap--topology is defined for a greater class of systems than the graph-topology. Zhu even proved that in general the class of systems, for which the graph---topology is defined, is a subset of the class of systems, for which the gap--topology is defined (see [18, p.853]). Because of these two reasons, we decided to work only with the gap-topology.

3. SUFFICIENT CONDITIONS FOR R O B U S T BIBO STABIITZ.ATION In this chapter we derive sufficient conditions for robustness of feedback stabilization. In the first section we study the relation between the gap-metric and Bezout factorizations. With the results of this section we are able to derive several bounds which guarantee the stability

of

a

perturbed

feedback

system,

if

the

perturbations

of

the

system

and

the

compensator are within these bounds. This is the content of the second section. 3.1. The gap--metric and Bezout factorizafions In this section we dig out, as said before, the relation between the gap-metric and Bezout factorizations. Many of the results presented here, will turn out to be useful in the next section. W e start with the following lemma.

LEMMA 3.1.1. Assume P ~ ~n*m(s), and regard P as an operator mapping a subspace of H~ m n to H~. Denote the graph of P by G(P). Let D ¢ B(I-I~) and N e B(H2,H2). Then:

(D,N) is an r.b.f, of P ¢=~ (1) G(P) = Range (2) Ker D = {0} PROOF "v" Suppose (D,N) is an r.b.f, of P. Because D is invertible, clearly we have Ker D = {0}. So we only have to prove (1). Suppose y e G(P). Then q x e Dom(P) c H~ : y = (x,Px) = (x,ND-lx). Because D is invertible, there exists a z e H~ such that x = Dz. This we can see as follows. Since (D,N) is an r.b.f, of P, there exist bounded operators Y and Z (Y and Z are matrices in M(~Ho~)) such that YN+ZD = I. Define z := (XP+Y)x. Then clearly z ~ H~ and we have: Dz = D(XP+Y)x = D(XND-I+YDD--1)x = D(XN+YD)D--lx = DD--lx = x So, indeed we have x = Dz and it follows immediately that:

y-,x

lx, _-

x ~ Dom(P) because Px -

z lDz, _-

Dz = N D q D z

zNz,

Ran 0 [ D ]

= Nz is bounded (N is bounded). So (x,Px) G

G(P) and we have y = (Dz, Nz) = (x,Px) ~ G(P).

22

"#" Now suppose (I)

and (2) hold. Assume (D1,N:) is an r.b.f, of P. Because

P ~ ~n*m(s) such a pair exists. By the first part of the proof we know that

G(P) = { (Dlz,N1z) I z e ~

}.

(3.2)

By (3.2) and (3.1) we know that for every x e H~ there exists an unique y e H~ such that

ix_- [o, ], and vice versa. Because (D1,N1) is an r.b.f, of P there exist boundeA operators Y,Z (Y and Z are matrices in M(IIIt )) such that YD t + ZN 1 = I. Therefore:

(YD + ZN)x = (Y,Z)[ D I x = (Y,Z)[ D: ]y = y. N1 So U

:=(YD + Z N ) m a p s H~ to ~

bijectively. Since [ D ] = r_] DNI1 a_] U, (D,N) must bc an

r.b.f, of P. o The next lemma is an ahcmative version of a result in [10, p.206].

LEMMA 3.1.2. Let P1,P2 e ~{n*m(s). Then I1(P1) maps G(P2) bijectively onto G(P:) if and only if 8(PpPz) < 1. In [9] Kato derived almost the same result (see [9, pp. 56-58, theorem 6.34]). His formuladon of the theorem, however, gives rise to the following surprising result.

LEMMA

3.1.3. Let PpP2 E ~n*m(s). If 8(P:,P2) < I, then

~(PI,P2) = ~(P2,P1) = 8(PI,P2). With help of lemma 3.1.1. and 3.1.2. we can now prove the following theorem.

THEOREM 3.1.4. Let P:,P2 e Rn*m(s), and (DI,NI) an r.b.L of Pp Def'me

Nl I"

(3.3)

23

Then: (D2,N2) is an r.b.L of P2

~(PI,P2) < I. PROOF "¢" Suppose 8(P1,P2) < 1. By lemma 3.1.2. we know that H(P 2) maps G(P 1) onto G(P2) bijectively. So

G(P9

-- I'I(P2)G(PI)=I'I(P2)Range [ DIN1] = = Range H(P2) [ N1 DI ] = Range [ D2 N2 I =

(3.4)

= { (D2z,N2z) ] z e H~ }. So I D a ]

maps H~ surjectively onto G(P2).

But we also have ker [ D~ ] = {0). This we can see as follows. Suppose [ ~

]z =

0. Then I'I(P2)[ DI N1 ] z = 0. And since I'I(P2) maps G(P1) bijcctively onto G(P2) we have

[°,] NI z = 0. So Dlz

= 0, and because (D1,N1) is an r.b.L we have, according to lemma

3.1.1., z = 0. We draw the conclusion that

[D]

[o2] N2 D

maps H~ bijectively onto G(P2). Now suppose

maps I~2 bijectively onto G(P2) S o

there exists

a

bijective U, such that [ D2 "~" Suppose (D2,N2) is an r.b.f, of Pz" Then we know by lemma 3.1.1.

G(P2)

= Range [ D2 N2 ] = Rang e ( FI(P2) [ D1 N1 ] ) = --

D1 ] : n(P2) (P0. n(P2)Range [ "1

So I-/(P2) maps G(PI) onto G(P2) surjectively. Now, let x e G(P1) and suppose FI(P2)x = 0. Since x e G(PI) and (DpNl) is an r.b.f. of P1, x e Range YI(P2)

N1

, so there exists a z e H~ such that x =

z = 0, and applying (3.3) we

m2

N1 z. So

z = 0. So D2z = 0 and because

(N2,D2) i s an r.b.f, of P2, this immediately yields z = 0. So x = 0, and II(Pz) maps G(P1) injectively onto G(P2). Finally we conclude that I'I(P2) maps G(P1) bijectively onto GfP2) and, according to

24 lemma 3.1.2. we have B(PpP2) < 1. [] It is also possible to derive an analogous result as theorem 3.1.4. for left-Bezout factorizations. We have the following theorem (for a proof we refer to [20, p.57]).

THEOREM

3.1.5. Suppose PPP2 e ~{n*m(s), and let (~i,~1) an l.b.f, of Pr Let ~1 and 131

denote the Hilbert-adjoint operators of lqI and 151 respectively. Define

(3.5) Then: ( N ~ , - ~ ) is an 1.b.f. of P2 ¢=~ 8(P1,P2) < I. Finally we mention the following result, which will be needed in the next chapter. With the tools developed in this section, it is not very difficult to derive this lemma. /.,EMMA 3.1.6. Let (DI,NI) and (D2,N2) be r.b.f.'s of PI and P2 e ~qn*m(s). Then: 8(PpP2) < 1 N1N2 + D1D2 is bijective.

[D2] is an r.b.f, of

PROOF According to theorem 3.1.4. 8(P1,P2) < 1 if and only if rI(P 1) N2

Pr From formula (2.16) and the fact that [ N1 D1 ] is an r.b.f, of P1, we know that

(D~,N~)" YI(PI) = [ D, N1 ](D~DI+N~N,)_I So YI(PI) [ N D22 ]--" [ D1 ](D~DI+N~N1)-1(D~D2+N~N2). N1 Because

[ D1N1]

I°21 is an r.b.f, of P1 if and only if

is already an r.b.f, of PI, YI(P1) N2

(DIDI+N1N 1) (D1D2+NIN2) is bijective. Since (DIDI+NIN1)1 is clearly bijective, we get as final result: 8(P1,P2) < 1 if and only if (DI*D2+N~N2) is bijective. []

25 3.2. Guaranteed bounds for robust stabilization In this section we present two bounds which guarantee the stability of a perturbed feedback system, if the perturbations of the system and controller are within these bounds. Suppose P o e

[{n'm(s) is the nominal system and Co ~ ~*n(s) the nominal controller,

and assume that P0 and CO form a stable pair, i.e. H(Po,Co) is stable. Let P~. ~ ~n*m(s) and C9. ~ [m*n(s) denote perturbed versions of Po and C O respectively. The In'st bound we derive with help of the tools developed in chapter 2. THEOREM 3.2.1. If H(Po,Co) is stable and

8(Px,P o) + 8(cx,c o) < I14 (I + II H(P0,Co) [12)-I/2

(3.6)

then H(P~,C~) is stable. PROOF From formula (2.12) it follows that

8(H(P~,C,L),H(Po,Co))_< 4 [8(Pk,Po) + 8(Cz,Co)]. So in this case we have

8(H(Pz,Cz),H(P0,Co)) < (1 + II H(Po,Co) 112)-v2. Now H(P0,Co) forms a stable closed loop system, so according to lemma 2.2.1. we know that H(P~,C~) is stable. [] To derive an other bound for robust stabilization, we need a few preliminary lemmas. /.,EMMA 3.2.2. Suppose A is an operator, mapping a Hilbert-space X onto X. Assume [[ A [I < I. Then (I+A) is invertible, and the inverse is bounded. PROOF Define

T :Then

~

(-1) k A k.

k=O

g[ V [I = Ilk_ ( - l ) ~ A k

It~

_ [i A Itk~ 1 - - It A I1"

26

So T is bounded and it is not difficult to prove that T = (I+A)-l:

(I+A)T = ~ (-l)k Ak+

~ (-l)k Ak+1=

k=O

k=O

k=l

T~I+A) -- ~: (-1)k Ak + k=0

k=l

~: (--1)k Ak÷~ = I. []

k=O

LEMMA 3.2.3. Suppose U o is an operator, mapping a Hilbert--space X onto X. Assume U 0 has a bounded inverse. Let U be an operator such that

II u - o o II < II u~l I[-~. Then U is also invertible, and the inverse is bounded. PROOF (Uo - U) = Uo(I - Us~U). If II w - u0 11 < 11 I - U 5 1 U

II -- II u~XWo - w 6 l u

So II u6au -

III

and

inverse is bounded.

that

the

l[

u6, I1"1, we have

II < II u61 II II U o -

u II < 1.

< 1. By lemma 3.2.2. we know that I + U~IU - I = Us1U is invertible, Since U o is a bounded

invertible operator,

it follows

immediately that U is also invertible, with bounded inverse, ra Now we return to our nominal system and controller Po and C o. Let (D0,N 0) an r.b.f. of Po and (lq0,lS0) an 1.b.f. of C o. Denote

Ao=

[ D°No],

Bo=

(I)o,So).

(3.7)

Define U o :-- B0Ao. We now have the following theorem (for a proof we refer to [15,

p.105D THEOREM 3.2.4. H(Po,Co) is stable if and only if U o is a bounded operator which maps H~ bijectively onto I-~2. With help of this last theorem and the foregoing lemmas , we can prove theorem 3.2.5.

27 THEOREM 3.2.5. Let (D0,N o) an r.b.L of Po and (lqo,l~io) an l.b.f, of C O, and assume H(Po,C o) is stable. Denote

w = [ II Ao II II BO II II u6ffi II ]. Then we have: if 5(P~,Po)

B(C'A,'Co)<

+

(3.8)

w"l

then

H(P~,C.~) is stable. PROOF First we show that the right-hand side of (3.8) is smaller or equal to one. We have I = BoAoU~I and so:

1 = II I II = II BoAoU61 II ~ II Bo II II Ao II II UO1 II = w. Thus w > 1, and w-1 < 1. According to (3.8) we know that 8(P~,Po) < 1 and ti(C~,Co) < 1. We can now apply the theorems 3.1.4. and 3.1.5. to get an r.b.f, of Pg and an 1.b.f. of C~.. Theorem 3.1.4. gives that (Dk,N~,) defined by [DT~ N~ ]

YI(P~.) [ D°N O]

=

is an r.b.f, of P~, and theorem 3.1.5. that (lq~j)~) defined by

NZ'

=

No ]

is an 1.b.f. of C~,. Analogous to (3.?) denote

N~.

B~. = (D~.,R~).

Then we have [I B~.A~.- BoA o II = = [I BxA~. - BoA~. + BoA~, - BoAo I[ = = I] (BL-Bo)A~, + Bo(A~.-Ao) ]I <

(3.9)

28 II A~. [I I1 B~,-Bo II ÷ II A~-Ao II II Bo II II A~, II Ilfn(C~)--n(co))So)l + II~(P~)--n(Po))Aoll II Bo II. And since A~. ffi I-I(P~.)Ao we get

II B~A~. - B o A o II <11 n(cx)-n(co) II + l[ n(P))-n(Po) II) II Ao II II Bo II = = II Ao II II so II (8(C~,Co) + 8(P~eo) ) < < I1"% II IlSo IIw-~= = II u ~ 11-I. Now according to lemma 3.2.3. U~ --- B~A~. is invertible and the inverse is bounded. By theorem 3.2.4. it is then clear that H(P~,C,A) is stable, n This last theorem will play a crucial role in the rest of this monograph.

4. OPTIMALLY ROBUST CONTROL In this chapter we tackle the problem of optimally robust control. Given a nominal plant P0, we want to design a compensator C that not only stabilizes P0, but also plants P in the neighborhood of Po- This neighborhood, stabilized by C, we want to make as large as possible. We

call

a compensator C O that stabilizes this largest neighborhood of P0 an

optimally robust controller. To find an optimally robust controller, we suggest the following approach. In theorem 3.2.5. we found a guaranteed bound w-t (w = [[ Ao II II B0 II II(BoA0)-lll) for robust stabilization. Now suppose P0 is the nominal plant, and assume that C is a controller that stabilizes Po. We also assume that the controller can be implemented exactly, so C.£ = C. Then theorem 3.2.5. becomes: if 8(P~,P 0) < w-l, then H(P~,,C) is stable, So all the plants in the ball around P0 with radius w-t are stabilized by C. Now w depends only upon Po and C, and P0 is fixed. So by changing C in S(P0) we can change the value of w-l. Now we are interested in a compensator CO e S(Po) for which w-t is as large as possible, because such a controller stabilizes all the plants in the largest ball around Po (the ball around P0 with radius w-l, where w-1 is as large as possible). In this way it is possible to find an, in a certain sense optimal, robust controller Co. In the second section of this chapter, we show that the approach suggested above, really leads to the solution of the problem of optimally robust control. Then, in the third section, we will derive a method to actually compute an optimally robust controller, and the radius of the ball around Po, stabilized by this controller. But before we can do this, we have to introduce an other new concept: the idea of normalized Bezout factorizations. This will be the content of the first section of this chapter.

4.1. Normalized Bezout factorizations Normalized

Bezout factorizations, we

mentioned

them

ah-eady in section 2.4., are a

special kind of Bezout factorizations. In this section we give a formal definition and show by construction that each plant P e M(~(s)) has a normalized right- and left-Bezout factorization. The state--space realizations of these factorizations will be needed in chapter 5, whcre wc develop an algorithm for the solution of the problem of optimally robust control for the case P e M(~p(S)). Vidyasagar, however, showed that also in the more general case, a plant P e M(~(s)) has a normalized right- and left-Bczout factorization (sec [15, p.262]). Suppose P E ~n*m(s), and let (D,N) be an r.b.f, of P. Define D*(s) := DT(-s) N*(s) := NT(--s). So D* and N ° are the adjoint operators of D operators from L 2 to I.2. Then we have the following definition.

and

and N respectively, seen as

3O

DEF.4.1.1.

Let P

~ Rn*m(s), We

say (D,N)

M(RI-I**) is

a

normalized right--Bezout

factorization (n.r.b.L) of P if: 1) (D,N) is an r.b.f, of P; 2) D*D + N*N = I.

(4.1)

Completely analogously we define: DEF.4.1.2. Let P ¢ Rn*m(s). We say (1~,~) e M(RH**) is a normalized left---Bezout factorization (n.l.b.f.) of P if: 1) (lq,13) is an 1.b.L of P; 2) lqlq* + B])* --- I.

(4.2)

Let R~*m(s) denote the set of all n*m matrices with elements in Rv(s). To prove for each plant P E R~*m(s) the existence of normalized right- and left-Bezout factorizations, we need the following theorem. It is a generalization of earlier results, stated in [12] and [16].

THEOREM 4.1.1. Let P ¢ ~*m(s), and suppose [A,B,C,D] is a minimal reaIization of P. Assume A ~ [ ]. Let X, respectively Y, be the unique positive definite solutions to the Algebraic Riccati Equations: (A-BH-1DTc)Tx + X(A-BH-1DTc) - XBH-1BTx + cTL-1C = 0

(4.3)

(A-BDTL-1C)y + Y(A-BDTLqc) T - ycTL-1CY + BH-1BT = 0

(4.4)

H :-- I + DTD,

where

L := I + DDT. A c := A - BF with F := H-I(DTc + BTx),

Define

A o := A - KC with K := (BDT + y c T ) L -1. Now define - Z o Yo

15oNo] :=

[ -HI/2F( s I-A o ) -1K "[L-1/2[I--C(sI-Ao)-IK]

Hltz[LI +F(s I -Ao)- 1(B--KI)) ] ] L-I/2[C(sI-Ao) -I(B-KD) +D]

and -No ~o

[-[(C-DF)(s I -Ac)-IB+DIH-112

Do 2o ] := [ [ I - F ( s I - A c ) - I B ] H - 1/2

J (4.5)

[I+(C-DF)(s I -A~) - 1KIL1/2 ] F(sI-A=)-IKL 1/2 (4.6)

31

Then (Do,N0) and (No,Do) are a normalized right-Bezout and a normalized left-Bezout factorization of P respectively. Moreover we have

PROOF (i) First we prove that all the matrices defined in (4.5) and (4.5) belong m M(IP&I ). To do this, it is su~cient to prove that the matrices A c and A o are stable, i.e. all the eigenvalues of the matrices A¢ and A o belong to the LHP. Suppose X is the unique positive definite solution m the ARE (4.3). Then we have: (A-BH-IDTc)Tx+x(A-BH-1DTC)-XBH-1BTx+cTL-1C---O. This equation can be written as

(a-BH-I(DTc÷BTx))Tx÷x(A-BH-I(DTc÷BTx))÷XBH-IBTx÷cTL-IC=0. Application of the definition of F yields (A-BF)Tx+x(A-BF)+XBHqBTx+cTLqG=0. And because A-BF = A~, we finally get ATX+XA c+XBH-1BTx+cTLqc---O.

(4.8)

Now we need the following auxiliary results: Claim 1: c T L q c = c T c - C T D H q D T c

(4.9)

Proof:

cTc-cTDHqDTc

= cT(I-DH-IDT)c = cT(I-D(I+DTD)qDT)c

cT(I-O+DDT)-IDDT)c

=

ffi cT((I+DDT)-I(I+DDT)-(I+DDT)-IDDT)c

=

cT(I+DDT)-IC = cTL-IC. Claim 2: XBHqBTx + cTL-IC = (C-DF)T(c-DF) + FTF Proof:

(C-DF)T(c-DF)+FTF

ffi cTc-cTDF-FTDTc+FTDTDF+FTF

ffi

(4.10)

32

cTc--CTDF-FTDTc+FT(I+DTD)F= cTc--cTDF_FTDTc+FTHF= cTc-cTDF-FTDTc+FT(DTc+BTx) = cTc-cTDF+FTBTx =

cTc--cTDHq(DTc+BTx)+(cTD+XB)H-IBTx= c T c - - c T D H - 1 D T c - c T D H - t B Tx+cTDH-1BTx+xBH-1BTx = cTL-1C+XBH-1BTx. From (4.8) and claim 2, we get:

A : X + XA c + (C-DF)T(C-DF) + FTF = 0.

(4.11)

Now suppose 2L ¢ c(Ac). Then there exists a vector x # 0, such that Aex = Lx. Pre--multiplying (4.11) with ~T and post--multiplying it with x, we get:

2 Re(X)(~Txx) + ~ T ( c - D F ) T ( c - D F ) x + ~TFTFx = 0.

We

now

(C-DF)T(c-DF)

prove and

FTF

that are

~T(C-DF)T(c---DF)x positive

+

~TFTFx

(4.12)

>

semi-definite. Assume that

0.

We

know

xT(c-DF)T(C-DF)x

that +

xTFTFx = 0. Then Fx = 0 and Cx =0. But %x = Aex = (A-BF)x = Ax, and Cx = 0. So is an unobservable eigenvalue of A, and this is in contradiction with the assumption that our realization [A,B,C,D] is minimal So we conclude: ~ T ( c - D F ) T ( c - D F ) x + ~TFTFx > 0. Now X is positive definite. So ~Txx > 0, and from (4.12) it follows immediately that Re(~.) < 0. So A c is a stable matrix. The proof that A o is a stable matrix is completely analogous. Suppose Y is the unique positive definite solution to the ARE (4.4). Then we have (analogous to (4.8)):

AoY + Y A : + ycTL-xCY + BH-IBT = 0.

(4.13)

Also we have the following facts (analogous to the claims 1 and 2): Fact 1:

BH-1BT = BB T - BDTL'qDB T.

(4.14)

Fact 2:

BH-1BT + y c T L ' q C Y = (B-KD)(B-KD) T + KK T.

(4.15)

From (4.13) and (4.15) we get:

AoY + Y A : + (B-KD)(B-KD) T + KK T = 0.

(4.16)

33

Now suppose ~. e t~(Ao). Then there exists a row vector y such that yA o

=

ky.

Pre-rnultiplying (4.16) with y and post-multiplying it with fiT we get: 2 Re(~.)(yYyT) + y(B-KD)(B-KD)Ty T + yKKTyT = 0. Completely

analogous

to

the

foregoing

case,

(4.17) we

can

prove

that

y(B-KD)(B-KD)TyT+yKKTy T > 0, because otherwise ~. would be an uncontrollable eigenvalue of A, which is in contradiction with the assumption that our realization [A,B,C,D] is minimal. Now Y is positive definite. So yy~T > 0, and from (4.17) it follows immediately that Re(~.) < 0. So A o is also a stable matrix. (ii) Secondly we prove formula (4.7). To do so, we f'trst mention the following fact: K C - B F = - ( A - K C ) + (A-BF) = A e - A o = (sI-Ao) - (sI-A e)

(4.18)

We now check the four equalities in (4.7). a) ZoN 0 + YoDo = I ZoN0 =

H1]2F(sI-Ao)-IK[(C-DF)(sI-Ac)-IB+D]H-1/2 = H112{F(sI_Ao)-IK(C_DF) (sI_Ac)-IB+F(sI_Ao)-IKD }H-112.

YoDo =

H1/2[I+F(sI-Ao)-1(B-KD)][I-F(sI-Ac)-1B]H-1/2 = H1/2 {I-F(sI-Ac)-IB+F(sI-Ao)-1(B-KD) -F(sI-Ao)-I(B -KD)F(sI-A ¢)-1B }H-I/2.

So, with help of formula (4.18) and the fact that K(C-DF) - (B-KD)F = KC - BF we get:

ZoNo+YoDo= = HI12{FfsI-Ao)-I(KC-BF) (sI-A c)-4B+I-F(sI_A~=CB+F(sI_Ao).4B }H'412 = HI/2 {F(sI-Ao)-~ [(sI-Ao)-(sI-A~] (sI_A~-IB÷I+ -F(sI_Ac)-IB+F(sI_Ao)-IB }H=II2 = HI/2 {F(sI-Ac)-IB-F(sI-Ao)-IB+I_F(sI_Ac)-IB÷F(sI_Ao)-IB }H-I/2 =

HI/2I H -1/2

=

I.

(4.19)

34

b) -Zo?o + Vo o = 0

-Zo? o =

-H I/2F(sI-Ao)-IK[I+( C-DF)(sI-Ac)-IK]L I/2 = -HI/2{ F(sI-A o)qK+F(sI-Ao)-IK(C-DF) (sI-Ac)qK }L %.

Voo=

H1/2[I +F(sI-Ao)-I(B-KD ) ]F(sI-Ac.)-IKL 1/2 =. H1/2 {F(sI-Ae)-IK+F(sI-Ao)-I(B-KD)F(sI-Ac)qK }L 1/2.

Again, formulae (4.18) and (4.19) give:

-Zo~o+Yo2o =

= H1/2{F(sI-Ac)-IK-F(sI-Ao)-qK+F(sI-Ao)q(BF-KC)(sI-Ac)qK}L 1/2 = =

H1/2{F(sI-Ac)-IK-F(sI-Ao)'-IK+F(sI-Ao)-I[(sI-Ae.)-(sI-Ao)](sI-Ac)-IK}L

= H1/2{F(sI-Ac)-IK-F(sI-Ao)qK+F(sI-Ao)qK-F(sI-Ac)qK}L

1/2 =

= H v2 o L v2 = O.

c) -~oNo + lqoDo = 0 -13oN o =

-L-I/2[I--C(sI-Ao)-IK] [(C-DF)(sI-Ac)qB +D]H q/2 = _L-1/2 {(C_DF)(sI_Ac)qB+D--C(sI_Ao)-IKD+ -C(sI-Ao)-IK(C_DF)(sI-Ac)-IB }I-I-1/2.

~0D0 =

L-I/2[C(sI-Ao)-I(B-KD)+D] [I-F(sI-Ar)-qB]H -1/2 =

L-I/2{C(si_Ao)-I(B_KD)+D_DF(sI_A ¢)qB+ --C(sI-Ao)-I(B-KD)F(sI-Ae)-IB}Hq/2. Canceling equal terms and using again (4.18) and (4.19), we get: -ISoNo+lqoDo= =

L-I/2{C(sI-Ao)-IB---C(sI-Ac)-IB+C(sI-Ao)-I(KC-BF)(sI-Ac)-IB}H -I/2 =

= L-l/2 {C(d-Ao)-IB-C(sI-A c)-lB + C(sI-Ao)-l[(sI-Ao)-(sI-Ac)](sI-Ar)qB }I-1-1/2 = _- L-I/2{C(sI_Ao)-IB_C(sI_Ac)-IB+C(sI-Ac)-IB--C(sI-Ao)-IBJH -1/2 = = L q/2 o I-1-1/2 = o.

I/2

35

d) 5 0 ~ 0 + ~0~0 = I L-1/2[I..-C(sI-Ao)-1K][I+(C-DF) (sI-A c)-tK]L :/2 = L-l/2 {I+(C---DF)(sI-A¢)-IK---C(sI-Ao)-IK+ ..C(sI_Ao)-IK(C--.DF)(sI_A e)-IK }L1/2.

~o~o

L-t/2[C( sI_Ao)-I (B -KD)+D] F ( s I - A c)-IKL1/2 =

=

I.,-1/2{C(sI-Ao)-t(B - K D ) F ( s I - A c)-tK+DF(sI-A c)-lK }L t/2. Application of (4.18) and (4.19) now gives:

~o%+1%Z0= = L-1/2{I+C(sI-Ac)qK-C(sI-Ao)-IK+C(sI-Ao)-I(BF-KC)(sI-Ao)qK}L1/2 = = L-l/2 {I+C(sI-Ae)qK-C(sI-Ao)-IK+ +C(sI-Ao)q[(sI-Ar)-(sI-Ao)](sI-A~)-IK }L l/z = = IfI/2{I+C(sI-Ac)qK--C(sI-Ao)-IK+C(sI-Ao)-IK--C(sI-Ac)-IK}L = L -I/2 1 L 1/2 =

(iii) Now

1/2 -.

I.

we

show

that

(D0,N 0) and

(1~o,13o) are

an

r.b.f,

and

an

1.b.f. of P

respectively. a) First we consider the right-Bezout case. The fact that (Do,N0) $ M(~J-I**) is already proven in i). So we only have to check the three conditions in definition 1.2.1.

1) Since D o = [I--FfsI-Ae)-tB]H-1/2, det(D o) tends to det(H q/2) as s

, -. Because det(H -1/2)

= det((I+DTD) -1/2) # 0, we know that D O is indeed non-singular. 2) The existence of two matrices Yo and Z o such that YoDo+ZoNo=I is already proven in fi.a). From i) we know that these matrices Yo and 7_.o belong to M(RH**). 3) So we only have to prove that P=NoD0 -1. Because D O is non-singular it is sufficient to prove that PDo=N o. To do so, we use the fact that BF

Then we have: PD o =

-- A

-

(A-BF)

=

A

-

A¢

=

( s I - A c) -

(sI-A).

(4.20)

36 = [C(sI-A)-IB+D][I-F(sI-Ac)-IB]H -'/2 = = {C(sI-A)-IB--C(sI-A)-IBF(sI-Ac)-tB+D-DF(sI-Ac)-aB}H -I/2 =

= {C(sI-A)-IB-C(sI-A)-I[(sI-A~)-(sI-A)](sI-Ae)-IB+D-DF(sI-Ac)-IB}H-1/2 = =

{C(sI-A)-iB-C(sI-A)-IB+C(sI-Ae)-IB+D-DF(sI-Ac)-IB}H-V2 =

=

{(C-DF)(sI-Ae)-IB+D}H-1/2 = N 0.

b) The left-Bezout case is completely analogous. In i) we have already proven that (~0,i30) e M(P,H**). Since 2 0 = L-I/2[I---C(sI-Ao)-1K], detO30) tends to det(L -1/2) as s

.- **, and because

det(L-1t2) = de~((I+DDT)-t/2) ~ 0, f30 is non-singular. The proof of the existence of two matrices 9 o and ~

in M(IP,H . ) such that IDo~o+~o,~o=I is already given in ii.d) and i). So

we only have to check the third condition of definition 1.2.2.: P=I30-11qo. Since ~30 is non-singular it is sufficient to prove that /)oP=~o. Now we use the following equality: KC = A - (A-KC) = A - A o ---- (sI-A o) - (sI-A).

(4.21)

Then we get:

BoP = = L-1/2[I--C(sI-Ao)-IK][C(sI-A)-IB+D] = = L-1/2{C(sI-A)-tB+D-C(sI-Ao)-IKC(sI-A)-IB-C(sI-Ao)-IKD} = = L-1/2{C(sI-A)-1B+D-C(sI-Ao)-tKD-C(sI-Ao)-t[(sI-Ao)-(sI-A)](sI-A)-IB} = = L-1/2{C(sI-A)-IB+D-C(sI-Ao)-IKD--C(sI-A)-IB+C(sI-Ao)-XB} = = L-II2{c(sI-Ao)-a(B-KD)+D}

= N o.

(iv) Finally we have to prove that both the Bezout factorizations are normalized, i.e. the equalities (4.1) and (4.2) hold. Again we start with the right-Bezout case. a) We are going to prove that D~D0+N~N0=I.

D~(s)D0(s) = DT(-s)Do(S) = = H-I/2[I+BT(sI+AT)qFT][I-F(sI-Ae)-IB]H-I/2 =

= H-t/2 {I-F(sI-A~)-:B+BT(sI+AT)-IFT-BT(sI+AT)dFTF(sI-Ao)-IB}H-1/2. (4.22)

37 N~(s)No(s) = NOT(-s)No(s) =

= H-I/2[-BT(sI+A:)-:(C-DF)T+DT][(C--DF)(sI-Ac)-iB+D]H -'112 = = H-I/2 {-BT(sI+A:)-:(C-DF)T(c-DF) (sI-A r)-:B-BT(sI-A:)-: (C-DF)TD+ DT(c-DF)(sI-Ae)-IB+DTD }I-1-112.

(4.23)

Adding (4.22) and (4.23), we get: D~(s)Do(s) + N~(s)No(s) = = I_I-1/2{_BT(sI+A:)q[(C_DF)T(c_DF)+FTF] (sI_A~)_aB+(I+DTD)+ BT(sI+A:) -1['FT-(c-DF)TD] +[DT(C-DF)-F] (sI-A e)qB }H -1/2.

(4.24)

Now we know that X is the unique positive definite solution to the ARE (4.3), and so equation (4.11) holds. Rearranging (4.11) gives: - ( A : X + XA c) = (C-DF)T(c-DF) + FTF.

(4.25)

Also we have the following fact: Claim 3: F T - (C-DF)TD = XB

(4.26)

Proof: FT-(c-DF)TD = FT-cTD+FTDTD = FT(I+DTD).--cTD = FTH.-CTD = = (cTD+xTB)H-IH---cTD = xTB = XB. Transposing (4.26) and multiplying it with -1, we get: DT(c-DF) - F = -BTx.

Application of the formulae (4.25), (4.26) and (4.27) in (4.24) finally gives: D~(s)Do(s) + N~(s)No(s) = = H-I/2 {BT(sI+A:)qiA:X+XAc] (sI-Ac)-IB+H+

BT(sI+A:)-IXB+_BTX(sI-Ac)-IB }H-I/2 = =

H-1/2{BT(sI+A:)-i[(sI+A:)X-X(sI-A r)] (sI-A ¢)-1B+H+ BT(sI+A:)-:XB_BTX(sI_Ac)-IB }I-I-1/2 _-

= H-V2{BTX(sI-A ~)-:B-BT(sI+AT)-:Xa+H+

(4.27)

38

BT(sI+AT)-IXB_BTX(sI-Ae)qB }Hq/2 _= ~ 1/2 rI Ia -V~ = I.

b) The left-Bezout case is again completely analogous: Ro(S)R~(s) = /~o(S)lqT(-s) =

= L-I/2[C(sI-Ao)-I(B-KD)+D][-(B-KD)T(sI+AT)-IcT+DT]L-I12 = =

L-V2{-C(sI-Ao)-t(B-KD)(B-KD)T(sI+AT)qcT+DDT+ C(sI-Ao)q(B-KI))DT-D(B-KD)T(sI+AT)qcT}L-I/2.

l~o(S)D~(s)

(4.28)

= Bo(S)13T(-s) =

L-1/2[I-C(sI-A )-aK] [I+KT(sI+AT)-acT]L-I/2 = L-I/2{I+KT(sI+AT)-IcT-c(sI-Ao)-IK--C(sI-Ao)-IKKT(sI+AT)-IcT]L-I/2. (4.29) Adding (4.28) and (4.29) gives: lqo(S)l~(s) + 5o(S)B~(s) =

= L-I/2{--C(sI-Ao)-I[(B-KD)(B-KD)T+IcKT](sI+AT)-IcT+cI+DDT)+ C(sI-Ao)-I[(B-KD)DT-K] +[KT-D(B-KD)T](sI+AT)-IcT }L q/2 .

(4.30)

Y is the unique positive definite solution to the ARE (4.4), and so (4.16) holds. From (4.16) we immediately get: -(AoY + YA T) = (B-KD)(B-KD) T + KKT.

(4.31)

Analogous to claim 3, and formula (4.27), we also have the following equalities:

(B-KD)DT - K = -YCT,

(4.32)

K T - D(B-KD) T = CY,

(4.33)

Using (4.31), (4.32) and (4.33) in (4.30), we find:

l~o(s)fl~(s) + Bo(S)B~(s) =

= L-V2 {C(sI-Ao)-I [AoY+YAT] (sI+AT)-IcT+L+

39

..-C(sI-Ao)-IycT+cY(sI+AT)-IcT}L -Vz __-L-If2{C(sI_Ao)-l[-(sI-Ao)Y+V(sI+AT)] (sI+AT)-tcT+L+ --C(sI-Ao)-,ycT+cY(sI+AT)-,cT} L-1/2= = L-I/2{-CY(sI+AT)-IcT+c(sI-Ao)-IycT+L+ -C(sI-Ao)--IycT+cY(sI+AT)-ICT }1-112= = 1. -1/2 L 1. -1/2 = I.

This completes the proof, n Theorem 4.1.1. gives a method for the construction of an n.r.b.f' and an n.l.b.f' for dynamical systems in M~p(S)). For computer purposes this method is implemented in the MATLAB-function ncoprfac. For the details of this implementation, we refer to appendix B. There is only one little problem left. For non--dynamical systems in M(Rp(S)), theorem 4.1.1. doesn't give a solution to the factorization problem, since for non--dynamical systems it is not possible to find a minimal realization [A,B,C,D] with A ;~ [ ]. But in this case we have the following theorem. THEOREM 4.1.2. Let P e ~'rn(s) and suppose P = K, with K a Define

I - Z Do 0

Yo ] lqo

:= [ -(I+KTK)-1/2KT

constant

transfermatrix.

(I+KTK)q/2 ]

(I+KKT)-I/2

(I+KKT )-I/2K

( I+KTK ) -1/2

KT(I+KKT)"t/2

(4.34)

and -No

~o

Then (D0,N0) and (lq0,~0) are a normalized right-Bezout factorization of P respectively. Moreover we have

i z0 g0] [ N0 13o l~o

Do

and a normalized left-Bezout

01 ~o

0

(4.36)

I

PROOF (i) Clearly all the matrices, defined in (4.34) and (4.35) belong to M(RI-.I**), because they are all constant matrices. (ii) Secondly we prove formula (4.36). We simply check the four equalities.

40

a) ZoYo+Y~o =

=

(I+KTK)'t/2KTK(I+KTK~-I/2+(I+KTK)-I/7(I+KTK)"t/2

=

= (I+KTK)-V2(KTK+I)(I+KTK)-V2 = L b) -Z0?o+Y0~ 0 = -- -(I+KTK)-I/2KT(I+KKT)-I/v+(I+KTK)"I/2KT(I+KKT) c) -DoNo+NoDo

-1/2 = 0.

=

= ---(I+KKT)'a/2K(I+KTK)-V2+(I+KKT)-I/2K(I+KTK)

"4/2 = 0.

d) DoVo+NoT.o = = (I+KKT)-I/2(I+KKT)-t/2+O+KKT)-I/2KKT(I+KKT)q/2

=

= (I+KKT)-I/2(I+KKT)(I+KKT) -I/2 = I. (iii) a) First we show that (Do,No) is an r.b.f, of P. We already mentioned that (Do,No) e M(~,I-I**). It is also quite clear that D O is non-singular, because D o = (I+KTK)-t/2 and (I+KTK) is a positive definite matrix. A solution to the Bezout-identity ZoNo+YoDo=I is given in ii.a), so we only have to prove that P = NOD51, and this is quite easy:

NoDSI = K(I+KTK)-I/2(I+KTK) I/2 = K = P. b) The left-Bezout case is completely analogous. Therefore we only give the proof of the third condition of definition 1.2.2.:

~i~11~0 = (I+KKTfl/2(I+KKT)q/2K

= K = P.

(iv) Finally we prove that both the Bezout factorizations are normalized. Again we start with the right-Bezout case.

a)

D~Do+N~N 0 = = (I+KTK)-I/2(I+KTK)-I/~+(I+KTK)q/2KTK(I+KTK)-t/2

=

= (I+KTK)-I/2(I+KTK)(I+KTK) q/2 = I. b) The left-Bezout case is completely analogous:

= (I+KKT)-I/2KKT(I+KKT)q/2+(I+KKT)-I/2(I+KKT)q/2 = = (I+KKT)'-I/2(KKT +I)(I+KKT)-I/2 = I.

41

This completes the proof, o We conclude this section with the following result, which is a generalization of the theorems 4.1.1. and 4.1.2. There we proved by construction the existence of an n.r.b.f, and an n.l.b.f, for a plant P G M(~p(S)) in both the dynamical and the non--dynamical ease. In [15, p.262] Vidyasagar also proved the existence of n.r.b.f.'s and n.l.b.f.'s for plants P e M(R(s)). This proof however is not constructive and is therefore not applicable for algorithmization. THEOREM 4.1.3. Each plant P ~ M(~(s)) has a normalized right-Bezout and a normalized left-Bezout factorization.

4.2. Optimally robust control In the introduction of this chapter, we suggested an approach to find an optimally robust controller. In this section we will describe this method more precisely and show that it really solves the problem of optimally robust control. Suppose P0 is the nominal plant, and assume that C ~ S(P0) is a controller that stabilizes P0. We also assume that the controller C can be implemented exactly, so C k -- C. With help of theorem 3.2.5. we can now calculate the stability radius w-1 of the controller C, i.e. the radius of the ball around P0, containing systems P, which are also stabilized by C:

w-1 = ( II A0 II II B0 II II (BoAo) -1 II )-1.

(4.37)

So when 8(P,P0) < w-l, then P is also stabilized by C. In the problem of optimally robust control, we want to design a compensator C, that not only stabilizes the nominal plant P0, but also plants in the neighborhood of P0- We want to make this neighborhood, stabilized by (2, as large as possible. Or, to put it more precise, we want to f'md a controller C with a stability radius as large as possible. Let wit denote this sharpest bound. In this section we will show that a controller C s that realizes this radius w~x is a solution to the problem of optimally robust control. But before we can do this, we first have to give a more formal definition of wil.

Let (Do,No) and Oqo,t$o) be an n.r.b.L and an n.l.b.f, of Po respectively. Let C o be one of the stabilizing controllers of Po and (Zo,Yo) and (~o,~o) an l.b.f, and an r.b.f, of C o respectively, such that

i z0 Y011 N0 o] ii 01 13o

lqo

Do

~o

0

I

(4.38)

42 Then we know from theorem 1.2.4. that the set of all stabilizing controllers is given by: S('Po) = { (Yo-RI~o)"I(Zo+R~0) { R e M(~-I ), IYo-R1qol . 0 } =: { (~0+DoR)(?o--NoR) -1 [ R ¢ M(IP,I-/**), [?o-NoR[ # 0 }. (4.39) From chapter I, we finally recall formula (I.II):

[ -Z°-R~°15o Y°-Rl~'° 11~/o

[ -N°Do ?o-NoR]~.o+Do R = [I0

0]i .

(4.40)

Now let C -- (Yo-R~o)q(Zo+R15o) be one of the controllers that stabilize P0- The inverse w of the stability radius wq belonging to C is then given by: w

-- [I Ao II [I B I] I[ (BAo)q [I-=

Now II

)I[ Do No ]

I°°l No

11 )] [CYo-R~o),(7-o+RDo)] II [I ([(Yo-R~o),(Zo+RDo)][ D° NO ])-1 11(4.41)

I1 -- 1 because (Do,No) is a normalized r.b.f, of Po and from (4.40) we see

that ((Yo-RNo),(Zo+I~o))

No

= I, so (4.41) becomes:

w = H [(Yo-RNo),(Zo+PdSo)l

)}.

(4.42)

With help of (4.42) we can now give a definition of the sharpest stability radius w~I. Let ~l~°n(s) denote the set of all stable mxn transfermatrices with elements in II(s): ~lm°n(s) :-- M(~-I ) ~ Utm*n(s).

(4.43)

Then we define wg :=

in f

[l

[(Yo-R~o),(Zo+R~o)]ll-

(4.44)

In section 4.3. we will show that this infimum is really achievable, and how a solution to Rg it, can be calculated. At present we simply assume that this is possible. We define for

43 E>O:

(4.45)

K(Po,e) := { P e ~a'm(s) I 8(P0'P) < E }.

From theorem 3.2.5. we know that, if Rg is a solution to (4.44), then Cg defined by

(4.46)

Cg := CYo-Rg~0)-I(Z0+Rg~0) stabilizes the whole set K(Po,w~t). W c

will now prove that such a controller Cg is a solution

to the problem of optirnally robust control by showing

that its stability radius Wzl is the

sharpest bound in the sense that thcrc are no compensators which can stabilize K(P0,e) with e > w~ z. In othcr words,

the largest numbcr

e, such that K(P0,~) can be stabilized by a

controllcr is w~ I.

First we define an other set, which also describes a neighborhood of the nominal plant PO" Let E > 0, then

R(Po.e ) := { P = (No+A.)(Do+Ad)-I e ~n*m(s) I

II [ % an ] tl <e }. (4.47)

where (D0,N0) is an n.r.b.f, of P0- Remark that R(P0,e) is independent of the choice of the n.r.b.f, of PoThe next theorem is a generalization of a result in [17, p.89].

THEOREM 4.2.1. Any controller C = (Yo-tLJqo)-I(Zo+t~o) stabilizes R(P0,e) if and only if

II [(Yo-R-~0),(Zo+R~5o)] II- ~t,

(4.48)

From this theorem and the definition of wg (formula (4.44)) we immediately derive: COROLLARY 4.2.2. The largest number e such that R(P0,e) can be controller is wg t, and if Rg is a solution to (4.44) then

stabilized by one

C~ = (Yo-Rg~0)-I(Z0+Rg~50) is a controller stabilizing R(P0,w~D. With help of some results, developed in chapter 3, we can now prove the following

44

crucial theorem. It is a generalization of an earlier result in [6], and proved this way by Zhu in [19, pp. 40-41]. THEOREM 4.2.3. If 0 < E ~ 1, then K(Po,e)

= R(Po,¢).

(4.49)

PROOF "~" Let P = (No+An)(Do+Aa)-t e R(Po,e), and II n.r.b.f, of Po, we have

II

No

II < E. Sinee (Do,No) is an

An

II = 1, and also [I (D;,N;) II = 2. Because

1

An j II < e <

2

we know by lemma 3.2.2. that (D;,N;) [ No+A Do+Ad n ] = I + (D~,N~)[

Ad A.]

is a bijective mapping. So, according to lemma 3.1.6., ~(Po,P) < 2. By lemma 3.1.3. we then know that g(Po,P) = ~(P,Po) = ~(Po,P) • Since (Do,No) is an n.r.b.f, of Po, the following statement holds:

11 [ DO No

]x 11=2

II x II = 2.

if and only if

With help of (4.50) we derive:

B(Po,P)

=

sup x e SG(Po)

=

sup x e H~, Ilxll=2

inf y e G(P) inf

l[ x - y

y ¢ G(P)

II

II =

[Do] No

x - y [I.

And because [ D0+AdN0+A n ]H~ ¢ G(P), wc find (notethat [ D0+AdN0+An I is bounded): ~(Po,P)

_< s up x ~ ~, llxIl=2

<

,, I

sup

x ~ ~,

inf l[ [ D° Ix - [ D0+Ad IY H y ~ H~ No N0+An

ilx[[=2

Ix- I

Ix ,,o

(4.50)

45

-- II [ Ad an ] II < eSo ~(Po,P) = ~(Po,P) < ~, and P e K(Po,O. "c" Let P e K(Po,e). Since 8(Po,P) < 1, we find, according to theorem 3.1.4., that (D,N') given by

No ] is an r.b.f, of P. And because (Do,No) is an n.r.b.f, of Po, we have:

N0 ] It= II -< II n ( P ) -

n(Po) II II

]

Do No

II

=

11 rI(P)

-I-I(Po)

II--

= 5('Po,P) < e. So P e R(P0,O. This complctes the proof. [] From theorem 4.2.3. and corollary 4.2.2. we finally derive the following solution to the problcrn of optimally robust control: COROLLARY 4.2.4. The largest number e such controller is w~1. If Rg is a solution to (4.44) then

that K(P0,E) can

bc

stabilized by one

C a = (Y0-RgI~0)-I(Zo+R~JSo) is a controller stabilizing K(P0,w~t).

4.3. Computation of an optimally robust controller and its stability radius In section 4.2. we reduced the problem of optimally robust control to the computation of an infimum, wg, and the construction of a matrix Rg, which realizes this bound. In this section we translate this problem to some well known mathematical problems. In the next chapter we will show how these problems can be solved when the plant P e M(~p(S)), and we wilI develop an algorithm that actually computes an optimally robust controller in this case.

46 At first, however, we start this section presenting the dual versions of the theorems 3.2.5. and 4.2.1. In the original versions of these theorems, we used a (normalized) r.b.f, of the plant Po and an l.b.f, of the controller. To derive the dual results, we simply have to change this order:, we use a (normalized) l.b.f, of Po and an r.b.f, of the controller C. So instead of the first parametrization of aU stabilizing continUers in formula (4.39), we take the second one:

C = (~,0+DoR)(~0-NoR)-L Let (Do,N o) and (~o,~o) be an n.r.b.f, and an n.l.b.f, of the nominal plant Po • ~n*m(s) respectively. Suppose C O is one of the stabilizing compensators of Po, and (Zo,Yo) and (~0,~0) are an l.b.f, and an r.b.f, of C O respectively, such that (4.38) holds. Then we have the following dual version of theorem 3.2.5. (Compare this result also with formula (4.42).)

THEOREM 4.3.1. Let R • ~,I-lm*n(s) and define

(4.51)

C O := (2o+DoR)(~o-NoR)-L

Assume P~, • ~n*m(s) and C~, • ~m*n(S) are perturbed versions of Po and C O respectively. Then the following statement holds: If

8(P~.,P0) + 8(c~,c0) < ( II [ ~~,o+Do ° - N ° RR 1 II )-i

(4.52)

then H(P~.C k) is stable.

Analogous to R(Po,e) we can define for e > 0:

T(Po,e) := { P -- (J50+~d)-l(J~0+~n) e ~n*m(s) ] 11 (~ct'7~n) l[ < e }. (4.53) Now we can formulate the dual version of theorem 4.2.1.

THEOREM 4.3.2. Any controller C = (~.o+DoR)(~o-NoR) -1 with R T(Po,e) if and only if

~

H~Hm*n(s) stabilizes

47

II [ 7°-N°R ~.o+DoR ] II ~ ~-t.

(4.54)

It is quite clear that the theorems 4.3.1. and 4.3.2. give rise to the following definition of fig, the inverse of the sharpest stability radius in the dual case:

~,s := inf R G ~I~°"(s)

ii ./ ~o_NoR r ~ [ 11 [ J ~ o + DRo

(4.55)

Let l~g be a solution to the infimum in (4.55). Then we define the dual optimally robust controller by

~g := (~.~0+D01~g)(~0-N01~g)-l.

(4.56)

Now fortunately we have the following theorem, which shows that both the primal and the dual case lead to the same set of solutions to the problem of optimally robust control. THEOREM 4.3.3. Define

v := D ~

- N~%.

(4.57a)

An alternative formula for V is then given by

V = Zol3~ - Yolq~,

(4.57b)

and for all R ~ Rl-I~*n(s) we have:

SI

[ ~o-NoR l

2o+DoR

II = tl [(Yo-Rl%),(Zo+RlSo)] II = ~/1 + 1[ V + R 112. (4.58)

PROOF First define Q :--

[5o lq°

, then o

[No l D o lq~

. Because (D0,N0) and

(lqo,13o) are an n.r.b.f, and an n.l.b.f, of Po respectively, Q is a unitary matrix on L~+n:

48 0 QQ* =

]~0 ~I0

DO lq~

=

-I3oNo+I~IoD 0

~o~+lqolq ~

=

0

Since multiplication with unitary matrices doesn't change the norms, we get: II [ ~?°-NOR Q[ ~°-N°R D~°-N~?° I IR II ~o+DoR ] II = H ~-o.DoR ] l[ = 11 [ I ] " [ 0 (4.59)

II [(Y0-RIq0),(Z0+RJ30)] II

=

II [(Zo+RI3o),(-Yo+RJqo)] II

=

= 11 [(Zo+RI3o),(-Yo+Rlqo)]Q* ]] = 11 [-I,(ZolS~-YOI~)] + R[0,I] 11.

(4.60) Now we have defined V := D~.o-N~P o. We will show that V = Zo~-Yo~ ~ too. Because Q[ ~0

~0

--

V

~o ~ o l I , , ] - - [ , ] ,

we have

i~o,o] : °'~I ~° ,o] °'Illv =

I

=

~o]iv DO 1~ -~°

}

NO }V Do

=

And with help of formula (4.38) we find: ]V o ,zo.[ ~o~ol o, zo-I -~° DO

+ (_Zo,Yo)[ 1~ "~ ]

=

= V + (-Zo~ ~ + Yolq~). So it follows that V = ZoEi~-Yol~~, and formula (4.57b) is proven. With help of this result, we can rewrite the formulae (4.59) and (4.60):

5;

49 1[ I T M I 1[ ~0+DoR

I]R[[" = I I [ v, 1+ [0

(4.61)

(4.62)

Ir [CYo-RI%),(Zo+R~o)] II = II [-I,V] + g[0,~ II.

We are now in a position to prove (4.58). Let R ~ ~P*n(s). Then we have:

x E =

sup x e a~.

( Ilxll 2 + II ( V + R ) x

llxll=l

Ilxll---x

sup

= 1 +

H~,

II2 ) --

II

x ~ H~,

(v+W)x

II2 = 1 +

II V+R

II2.

(4.63)

Ilxll--1

And also

11 [A,v] + R[0,I] II2 = l{ [-I,(V+R)] 112 -= sup ( Ilxll 2 + II ( V + R ) x x ~ H~,

-- 1 +

II2) =

IIxll--1

sup x E H~,

II ( V + R ) x

112 = 1 + II V + R

II2.

(4.64)

Ilxll=l

From (4.61)--(4.64) it follows immediately that for each R ~ RHm'n(s): [ ~°-N°R ]

II

2o+DoR

II = J_ 1 + II V+R II2 = II [fVo-gl~o),(Zo+Vd3o)] II

and this is exactly formula (4.58). This completes the proof. [] As a direct consequence of theorem 4.3.3. we have that both the characterizations of the sharpest stability radius, w~l in the primal case (formula (4.44)) and ~,~1 in the dual case (formula (4.55)) are the same, i.e. w~l = ~,~1.

(4.65)

Moreover, if R s is a solution to the infimum in (4.44) then R s is also a solution to the infimum in (4.55). And also the other way round: if l~s is a solution to the infimum in (4.55), then I~8 is a solution to the infimum in (4.44) too. So by theorem 1.2.5. we know

50

that the optimally robust controller in the primal case, Cg (formula (4.46)) is equal to the optimally robust controller ~s

(formula (4.56)) in the dual case. So the primal and dual

method deliver the same solutions to the problem of optimally robust control. Finally, with help of theorem 4.3.3. we are able to actually compute wg and Rg. Combination of the formulae (4.44) and (4.58) gives:

Wg

= /1

= + (

i nf ~/1 + II V+R II2 = / 1 R ~ ~rq~*~(s) i nf II V÷R II )z. R e ~-~*"(s)

+

i nf [1 V+R 112 = R E Er-/~*.(s) (4.66)

SO, m determine the stability radius W~1, and a matrix Rg which realizes this bound, we have to compute ot :=

i nf II V + R II R ~ ~LI~*n(s)

(4.67)

and a solution R 8 to this infimum. But this is a well known mathematical problem: a so called standard Nehari-problera. And the value of ~ can be calculated as the norm of a certain

Hankel-operator. In the next chapter we show how these mathematical problems can

be solved when P ~ M(~p(S)).

We are then able to develop an algorithm that solves the

problem of optimally robust control in this case.

51 5. AN ALGORrIS{M TO COMPUTE A N o F r I M A I I Y

ROBUST CONTROI.I.I~-R AND ITS

STABILITY RADIUS In chapter 4 we translated the problem of optimally robust control to some well known mathematical problems. In this chapter we show how these problems can be solved. Finally we derive an algorithm that computes

an optimally robust controller for each plant P

M(Rp(s)). The stability radius, belonging to this compensator, is also calculated. Note that we confine ourselves to the case P ~ M(~p(S)). This will be a standing assumption during the last chapters of this book. In the first section of this chapter we treat the non-dynamical case. This case is very simple, and computation of an optimally robust controller can be done by hand. In the rest of the chapter we deal with the dynamical case. First, in the second section, we figure out how a state--space realization of the the matrix V (formula (4.57)) can be calculated. With this knowledge we are able to compute the value of the infimum a

(formula (4.67)) of

section 4.3. Here the concept of I-Iankel--operators turns out to be very useful. In the fourth section we give a solution to the so called Nehari--problem

i nf

R ~ ~-I~*n(s)

II V+R {I; we

find a matrix R that realizes the value of this infimum. At the end of this chapter, in section 5, we finally put all the pieces together and describe an algorithm that computes an optimally robust controller and its stability radius in the dynamical case. 5.1. The non-dynamical ease As already mentioned in the introduction of this chapter, we treat the non--dynamical case separately. The main reason to do so, was given in section 4.1. When the transfermatrix of a plant P E ~*m(s) is a constant matrix, we can't use the method of theorem 4.1.1. to construct a normalized r.b.f, and 1.b.f. of P. Instead we have to use the construction of theorem 4.1.2. Fortunately this leads to a very simple solution to the problem of optimally robust control. In section 4.3. we saw that in the end the problem of optimally robust control can be reduced to the computation of the infimum ct :=

inf

R ~ ~-Im*n(s)

(where V = D ~ o - N ~

I[ V+R II

o = ZoI3~-Yo~ ~ ) and a solution R to it. We will first solve this

problem, and with the solution of it, compute the optimally robust controller and the stability radius belonging to it. Let P0 ~ ~*m(s) and suppose Po = K, with K a constant transfermatrix. With help of the formulae (4.34) and (4.35), which describe an n.r.b.f,

and an n.l.b.f, of P0, we can

52 calculate V:

V = (I+KTK)-t/2KT(I+KKT) -1/2 - (I+KTK)-I/2KT(I+KKT) -1/2 =- 0.

(5.1) So, we get oc =

i nf [[ V+R [I = 0, and Rg = 0 is a matrix which realizes this R e ItHm'~(s)

bound. With help of formula (4.46) and again (4.34) and (4.35) we find the optimally robust controller Cg:

c s = (Yo-RgI%)-I(Zo+R,Do) = Y~IZo = _- (I+KTK)I/2fI+KTK)-I/~KT = K T.

(5.2)

The stability radius w~l can be calculated with formula (4.66):

w~ t = ( 1 + (

i nf II V+R II )2 )--I/2 = (1 + 02) -1/2 = 1. R e RHm***n(s)

(5.3) So, we have proven the following theorem:

THEOREM 5.1.1. Let P e ~*m(s) and suppose P = K, with K a c o n s t a n t transfermatrix. An optimally robust controller Cg stabilizing P is then given by

Cg = K T and the stability radius belonging to Cg is

w~1=

I.

So C z stabilizesK(P,I).

5.2. A state---space realiTation of the matrix V In this section we start with the treatment of the dynamical case. First we have to find a way to compute the matrix V, or to put it more precisely, we have to construct a state--space realization of the matrix V. It is of course possible to substitute the state--space realizations

of

theorem

4.1.1.

into

formula

(4.57)

and

then

to

construct

a

state--space

53

realization of V with help of the multiplication and addition formulae given on page ix. But this gives a solution of a very high order. The next theorem shows that there is a much better way to find a state--space realization of V.

THEOREM Assume

5.2.1. Let P e P~*"a(s), and suppose [A,B,C,D] A

~

[ ]. Let X, respectively Y

is a minimal realization of P.

be the unique positive definite solutions to the

Algebraic Riccati Equations (4.3) and (4.4). Define as in theorem 4.1.1.:

A c := A - B F with F := H-1(DTc+BTx)

(where H = ( I + D T D ) )

A o := A - K C with K := (BDT+ycT)L -:

(where L = (I+DD T) ).

Then we have the following two state-space formulae for the transfermatrix V, defined in (4.57):

PROOF

V = -H"i]2DTL 1/2 + H-1/2BT(sI+A:)-I(I+XY)CTL-V2 =

(5.4)

= _HI]2DTL-1]2 + H-1]2BT(I+XY)(sI+A:)-IcTL-1/2.

(5.5)

With help of the solutions X and Y to the ARE's (4.3) and (4.4), we constructed in

theorem 4.1.1. an n.r.b.L (D0,N o) and an n.l.b.f. (lq0,l)o) of P, and matrices Yo,Zo,~0 and ~o such that (4.7) holds. W e here simply substitute these formulae in the definition of V (formula (4.57a)), and calculate a state--space realization of V without using the multiplication and addition formulae of page ix. In this way we get:

V = D~. o - N~ o with

D~to

= DT(~)~(s)

=

= H-V2[I+BT(sI+A:)-:FT]F(sI-Ac)-IKLV2

=

= H-l/2 {F(sI-A c)-IK+BT (sI+A:)-IFTF(sI_A c)-tK }h 1/2,

=

-Noz

-:o(s)

(5.6)

=

= I-I-]/2[BT(sI+A:)-I(C_DF)T-DT][I+(C_DF)(sI_Ac)-IK]L]/2 = =

- D T - D T ( c - D F ) (sI-Ac)-lK }L 1/2. Because X is a solution to the ARE (4.3), we know by (4.11) that

(5.7)

54

(C-DF)T(c---DF)+FTF = -(ATX+XA¢)

= -(sI+AT)x+X(sI-A¢).

(5.8)

We also recall formula (4.27): F - DT(c-DF) = BTx.

(5.9)

Application of (5.8) and (5.9) in the addition of (5.6) and (5.7) gives

v

= D#o

-

N~Vo =

= H-lh{ BT(sI+AT)q[(C-DF)T(c-DF)+FTF] (sI-Ac)-IK+ +BT(sI+AT)-](C-DF)T+[F-DT(C-DF)] (sI-Ac)-IK-DT} L :h = =

H-1/2 {BT(sI+AT)-I[.--(sI+AT)x+X(sI-A ¢)](sI-A c)-IK+ +BT(sI+AT)-I(C-DF)T+BTx (sI-A c)--lK-DT }L 1/2 =

= H-V2 {-BTx(sI-Ac)-]K+BT(sI+AT)-]XK+BT(sI+AT)-I(C-DF)T+ +BTX(sI-Ac)-IK-DT}L 1/2 = = H-1/2{-DT+BT(sI+AT)-I[XK+(C-DF)T]}L 1/2.

(5.10)

Now we need the following auxiliary result: Claim 1: [XK+(C--DF) T.JL 1/2 = (I+XY)CTL 1/2.

(5.11)

Proof: [XK+(C-DF)T]L1/2 = {X(BDT+ycT)L-I+cT-FTDT }L]/2 = = {(xBDT+x-ycT)L-I+cT--cXB+CTD)H-1DT}L

112 =

= {(XBDT+xYcT)L-I+cT--(XB+cTD)DTL-1}L

1/2 -_

= {xBDTL-I+xYcTL-q+cT-xBDTLq--K:TDDTLq}L1/2 = = {xYcTLq+cT((I+DDT)_DDT)L_I}LV2 = (xYcTL_t+cTLq}L1/2 = _- (I+Xy)cTL-I/2, where we used the fact that H-ID T = (I+DTD)-ID T = DT(I+DDT)-I = DTL-I. Now by substitution of (5.11) in (5.10), we immediately get:

(5.12)

5S

V -- -I-Iq/2DTL I/2 + I-I-1/2BT(sI+A:)-I(I+XY)CTL -1/2 and this is exactly formula (5.4). To prove (5.5), we use the other characterization of V as given in formula (4.57b): v

ZoO:

=

z ~

zo(,)OoT(-s)

=

-

Yol~.

=

_- HV2F(sI_Ao)-:K[I+KT(sI+A:)qcT]Lq/2 =

= HV2{F(sI-Ao)-:K+F(sI-Ao)-:KKT(sI+A:)qcT}LV2` -Y?;

=

Yo(s)~oT(-~)

(5.13)

_-

= HI/2[I+F(sI-Ao)-:(B-KD)][(B-KD)T(sI+A:)-IcT-DT]L q/2 =

= H:I2{F(sI-Ao)-I(B-KD)(B-KD)T(sI+A:)-IcT-DT+ +(B-KD)T(sI+A:)-ICT-F(sI-Ao)-I(B-KD)DT }L -I/2.

(5.14)

Since Y is a solution to the ARE (4.4), we know, according to (4.16): (B-KD)(B-KD)T+KK T = - A o Y - Y A : = (sI-Ao)Y-Y(sI+A:).

(5.15)

We also recall formula (4.32) K - (B-KD)D T = YC T. With help of (5.15) and (5.16), addition of (5.13) and (5.14) yields

v = zo~

- YoS; =

= H1/2{F(sI-Ao)-1[(B-KD)(B-KD)T+KKT] (sI+A:)qcT-DT+ +F(sI-Ao)-I[K-(B-KD)DT] +(B-KD)T(sI+A:)-IcT} L -I/2 = H1/2{F(sI-Ao)-I [(sI-Ao)y-y(sI+A:)] (sI+A:)qcT-DT+

+F(sI-Ao)-IycT+(B_KD)T(sI+A:)-qC T }L-I/2 -_ = HII2{FY(sI+A:)-IcT-F(sI-A o)-IycT-DT+ +F(sI-Ao)-IycT+(B-KD)T(sI+A:)-IcT }L q12 =

(5.16)

66

= HV2{-DT+[FY+(B-KD)T](sI+A:)-IcT}L -112.

(5.17)

Finally we have to use the following result: Claim 2: H]/2[FY+(B-KD) T] ffi

'/,Bra+xv)

(5.18)

Proof: With help of formula (5.12) we find: H1/2[FY+(B-KD) T] = HI/2{H-10DTc+BTx)Y+BT-DTK T} = = HI/2{H-IDTCy+H-IBTXY+BT-DTL-I(DBT+cyT)}

-_

= H1/2{H-tDTcY+H-tBTxY+BT-H-]DTDBT-H-1DTcy} =

=

Ha/2{H-1BTxy+(H-I(I+DTD)-H-tDTD)B T} = Ht/2{H-1BTxY+H-1B T} _-

= H-1/2BT(I+XY). Substituting (5.18) into (5.17), we get: V = - H I / 2 D T L -1/2 + H-I/2BT(I+XY)(sI+A:)-tcTL -1/2 and this is exactly formula (5.5). This completes the proof, because in theorem 4.3.3. the equality of the two formulae for V was already shown, o Although not necessary, it is also possible to prove the equivalence of (5,4) and (5.5) directly. First of all application of (5.12) yields: I-I-1/~DTLI/2 =

H-I/2I-IH-1DTLi/2 =

HI/2DTL-1L1/2 = HI/2DTL-:/2.

So to prove the equality of (5.4) and (5.5) it is necessary and sufficient to prove that: (sI+A:)-I(I+XY) = (I+XY)(sI+A:)-I. But this equality can also be simplified:

(sZ+AT)-t0+XY) = O+XY)(sI+AT)-~

(I+XY)(sI+A:) -- (sI+A:)(I+XY)

(5.19)

57

(5.20)

(I+XY)AoT = AT(I+XY ). Now we have:

(I+XY)ATo = (I+XY)(AT-cTKT) = (I+XY)(AT--cTLq(DBT+cy)) = = AT-cTL-IDBT-cTL-,CY+XYAT_xYcTL-,DBT_xYcTL-1Cy

(5.21)

and

AT(I+XY) = (AT-FTBT)(I+XY) = (AT-(cTD+XB)H-,BT)(I+XY) = = AT_CTDH_IBT_XBH_IBT+ATXy--CTDH-IBTxY-XBHqBTxy.

(5.22)

Since DH-1 = L I D , combination of (5.20), (5.21) and (5.22) gives:

(I+XY)AoT = AT(I+XY) -cTLq CY+xYAT-xYcTL-~DBT-xYcTLqcY =

=

-XB H-IBT+ATxy--cTDH-IBTxY-XB H-IBTx y

X Y (AT_.cTL_IDB T)_xYcT L qCY +XBH-ZB T -= (AT---cTDH-IBT)xY-XBH-IBTxY+cTL-ICy X {y ( A T - c T L - I D B T ) - y c T L - I C Y + B H - I B T } -= {( A T - C T D H q B T ) x - X B H - I B T x + c T L - I C }y

(5.23)

With help of the APE's (4.3) and (4.4) we can transform (5.23) to: (I+XY)AT _-- AT(I+XY)

X(-(A-BDTLqc)y) -_ (-X(A-BH-,DTc))y. And using the fact that DTL -1 = H-aDT (formula (5.12)), this last equation is easy to prove:

X(-(A-BDTL-IC)Y) = -XAY+XBDTL-ICy = - X A Y + X B H q D T c y =

58

= (-X(A-BH-tDTc))y. So this completes the proof of formula (5.20) and therefore also the direct proof of the equality of (5.4) and (5.5). We remark here that the state--space realizations of V given in (5.4) and (5.5) have the same order as the original plant P. But this is not the only property of these formulae. If [A,B,C,D] is a minimal realization of P, the given state-space formulae of V are als0 minimal. This result is stated in the next theorem.

THEOREM 5.2.2. Let P e ~*m(s) and suppose [A,B,C,D] is a minimal realization of P. Assume A # [ ]. Then the state-space realizations of the matrix V, associated with P as defined in (4.57), and given by the formulae (5.4) and (5.5): V(s)

t-AT,(I+XY)CTL-V2~-II/2BT,-H-V~DTL I/2]

(5.24a)

= [_A:,cTL-I/2~-I-I/2B T(I+XY),_H 1/21)TL-I/2]

(5.24b)

=

are also minimal. PROOF a) We start with the proof of the minimality of the realization (5.24a). i) First we show that the pair (-A:,(I+XY)CTL -1/2) is controllable. Because L -1/2 is invertible it is enough to prove that (A:,(I+XY)CT) is controllable. After transposition this is equivalent with the claim that (C(I+YX),Ae) is observable. And this last statement is easy to prove with help of formula (5.20). Becaus~ X and Y are the solutions of the ARE's (4.3) and (4.4) we have (5.20):

(I+XY)A: = A:(I+XY). From this we immediately get (after transposition):

Ao(I+YX) = (I+YX)A¢.

(5.25)

Since (I+YX) is positive definite and therefore invertible, we have A c = (I+YX)qAo(I+YX). With (5.26) we derive that (C(I+YX),Ac) (C(I+YX),fI+YX)-tAo(I+YX)) is observable. Because

(5.26) is observable if and only if we can consider (I+YX) as a basis

59 transformation, this is equivalent with (C,Ao) is observable. So, since A o = A-KC, we have to prove that (C,A-KC) is observable. But because [A,B,C,D] is a minimal realization of P, this last statement is immediately clear from the fact that (C,A) is observable. ii) The second part of the proof is quite easy.

(H-I/2BT,-AT)

is observable if and only

if (Ac,BH-1/2) is controllable. Since H -~/2 is invertible this is equivalent with

(Ac,B) is

controllable. And this last statement follows immediately from the definition Ac = A - B F and the fact that (A,B) is controllable. So (A-BF,B) is clearly controllable. b) The proof of the minimality of the realization (5.24b) of V is completely analogous, and is therefore omitted. [] For computer purposes the construction of this state-space realization of V has been implemented in the MATLAB-funetion speeeopr. More information can be found in appendix B.

To

conclude

-H~/2DTLt/2, the matrices Ae

and

this

section

we

finally remark

that,

apart

from

the

constant

term

matrix V is completely anti-stable. In theorem 4.1.1. we proved that the A o are

stable matrices,

so - A T

and

-A T

are

anti--stable: all their

eigenvalues belong to the RHP. From this fact we derive that in the dynamical case the stability radius w~t is always smaller than one. Because V is anti--stable we know that

e =

i nf II V+R II > 0. R • ~"I~*n(s)

And with help of this inequality we get

wg = ~/1 + (

i nf II V+R II )z > 1, R E RHm*n(s)

SO

w~l = 1/w s < 1.

(5.27)

This result is not very surprising. Since we know that the gap between two plants is always smaller or equal to one, we can't expect to find an optimally robust controller with a stability radius equal to one in the dynamical case. The exact calculation of the stability radius w~ 1 however, will be the subject of the next section.

5.3. Computation of the optimal stability radius w~t; Hankel--operators With help of the state--space realization of the matrix V, found in the last section, we

60 are now able to derive a method to calculate the sharpest stability radius w~ 1 for each plant P e M(~p(s)), as defined in (4.44). The infimum we want to calculate, turns out to be equal to the norm of a certain Hankel--operator. First of all, however, we have to give a few more definitions. (In this section we often drop the superscripts for notational convenience. The dimensions are then clear from the context.)

DEF.5.3.1. The space L~ consists of all functions x : i~

) (n, which are squarc-(Lcbesgue)

integrable. L~ is a Hilbert-space under the inner-product

(x,y)

:=

(2~) "-1

" " J ~ x(lo)*y0c0)dc0.

(5.28)

We see that H 2 is a closed subspace of L 2. Of course its orthogonal complement H~ is closed too. So it is possible to introduce the following orthogonal projections: FIp the orthogonal projection from L 2 onto H i , and 1"I2, the orthogonal projection from L 2 onto H 2. DEF.5.3.2. The space L** consists of all functions F : i~ ) C for which I F(ic0) I is essentially bounded (bounded except possibly on a set of measure zero). The L**-norm of F is defined as [[ F I1,, := e s s sup I F(ico) l"

(5.29)

With this norm, L** is a Banach-space. DEF.5.3.3. P~L, := ~(s) r~ L**. So, R L

is the subspace of L** which consists of all real-rational proper functions

without poles on the imaginary axis. Let M(L**) and M(RL**) denote the sets of all matrices with elements in L** and P,L** respectively. A matrix F e M(L**) can then be seen as an operator mapping L 2 into I..2: FL 2 c L.2. We can now give the following definition: DEF.5.3.4. Let F e MfL**). The Laurent-operator with symbol F, denoted AF, is an operator from 1.,2 to L 2 and is defined as AFg := Fg II A F II -- II F I1~,, so A F is bounded. We are now ready to introduce the Hankel-operator with symbol F.

(5.30)

61

DEF.5.3.5. Let F ¢ M(L®). The Hankel--operator with symbol F, denoted I F , maps H 2 to H i, and is defined as

FF := nlhF I r h.

(5.31)

Remark that when F e M(H**), then F F = 0 because FH 2 c H 2, so the projection onto H i, H l, maps this whole set to zero. The relation between Hankel--operators and the infimum we want to calculate is given by Nehari's theorem: THEOREM 5.3.1. Let V ~ M(RL**). Then we have:

i)

in f II V - R R E M(~H,,)

(5.32)

I1,0 = II r"v II

ii) There exists a matrix R ~ M(YJ-I**) such that [I V - R II = II 1"V IIPROOF This theorem was stated in a slightly different way by Francis in [4, p.59]. For the proof however, we refer to [13]. How to find an R such that II V - R the subject of the next section.

I1 = II r v

II will be

With this theorem we are able to rewrite the formula for wg, the inverse of the maximal stability radius.

we= 4/ 1 + (

i nf

R ~ ~a-I~*n(s)

II V+R II )2 = ~/1 + II r y II2.

(5.33)

So, to compute the maximal stability radius we only have to find a method to calculate the norm of the Hankel-operator with symbol V, when a state-space realization of V is given. In the rest of this section we describe such a method. For a complete proof of it, and more details, we refer to [4, eh.5]. Let V ~ M(RL**). There exists a unique decomposition (for example by partial-fraction decomposition) V = V 1 + V 2, where V 1 is strictly proper and analytic in Re(s) < 0, and V2 is proper and analytic in Re(s) > 0, i.e. V 2 ~ M(~*-I**). So F V --- FV1 + FV2 = FV1 because F V 2 = 0. Introduce a (minimal) realization of VI:

Vl(S ) -- [A,B,C,0].

(5.34)

Then the matrix A is anti--stabIe. In the case we are interested in, with V as in formula

62

(5.4), this decomposition is quite simple. V 2 consists only of the constant matrix ---H-I/2DTLI/2 and a realization of V I is given by T + T-I/2 -I/2 T Vl(S) = [ - A c , ( I XY)C L ,H B ,01.

(5.35)

W e now return to the general case, with V 1 defined as in (5.34) and A anti-stable. We define two auxiliary operators, the controllability operator ~e: Vc : L2[0,~*)

, •n :

~¢u := J** e-AV'Bu(x)d%

(5.36)

and the observability operator ~o:

gto : £n

, L2(--~,0] :

(~oX)(t) := ceAtx

(t < 0).

(5.37)

By the inverse Laplace--transformation we can calculate the time-domain analogue rVl of the Hankel--operator FVt. This operator turns out to be equal to:

r~ 1

=

VoVo.

(5.38)

(For a proof, see [4, p.54].) Since the Laplace-transformation is an isometry, it preserves the norms of the operators. From this fact the equality II rvl II -- II rvl II follows immediately. We also need a little more operator theory. Let 0 be an operator from a I-Iilbert-space X to a Hilbert--space Y. Let 0* denote the adjoint of ~. Then ~*0 is self-adjoint and we have

II ~ II2 = II 0"~ II (for I1

a

proof,

see

[11,

rvl 11--/11 r~qrv~

I1-

(5.39)

p.199]).

So,

From

instead

this

equation

of computing

we

[I FV1

immediately II, we

can

derive also

that

calculate

II r~,xrva II, and this can be done with help of the next theorem. THEOREM

5.3.2. The

spectrum t~fF~lrV1 ) of F~rlFV1 is not empty. The eigenvalues of

r~rv~ a~

real and nonnegative and the largest of them equals II

r~rv~ II.

PROOF A proof of the first claim is given in [4, p.57]. The rest of the proof can be found in [11, sec.9.1-9.3].

63 We now define the controllability and observabitity gramians:

L c :=J**e-AtBBTe-ATtdt

(5.40)

L o :=0 J**c--ATtcTce-A tdt.

(5.41)

It is not difficult to show that L e and

Lo

axe the unique solutions of the Lyapunov

equations: A L e + L~A T

BB T

(5.42)

ATLo + LoA = c T c .

(5.43)

=

It is also quite easy to prove that the matrix representations of ~c~e and VoVo are L c and Lo respectively. This last fact gives rise to the following theorem. THEOREM

5.3.3.

The

operator

F~/1FV1

and

the

matrix

LcL o have

the

same

nonzero

cigenvalues. PROOF i) Let ~. ~ 0 be an eigenvalue of F~tFv1. Then obviously ~. is an eigenvalue of F*Fvlvl' the time-domain analogue of F~/1FVt, which equals ~Vo*~oVe" So there exists a u ~ 0 in I.,2[0,**) such that

VcVo¥oVeU = 7tu.

(5.44)

Pre---multiplying (5.44) by Vc and defining x := Veu, we get: L~LoX =

* ,

=

=

Clearly x ~ 0, because otherwise Veu were to equal zero, and so would 2~u from (5.44). Since both ~. and u arc nonzero, this is not possible and ~. is an cigenvalue of LcI..o. ii) Let ~ ~ 0 be an eigenvalue of LcL o, and x ~ 0 a corresponding eigenvector, so LcLox = 2Lx.

(5.45)

Pre---multiplying (5.45) by v~L o and defining u := ~ L* oX , we derive: *

*

*

*

*

x?c~o~o¥~u . VcVo~oVeVcLox . .

*

V~LoLcLox .

v~Lo)tx = ~.VcLox *

~.u.

64

From (5.45) it also follows that Xx = L~Lox = ~g¢~g~L,ox = ~gcu. Because X and x are both nonzero, Vcu # 0, and this is only possible when u # 0. Therefore X is an eigenvalue of Vc~goV/oV/c, and hence of F{/ 1Fv1. r~ • • Combination of theorem 5.3.2. and 5.3,3. yields: COROLLARY 5.3.4. ]1 F~IFV1 [I equals the largest eigenvalue of LcL o.

So we have found a way to calculate II FV1 [I --- ,,/11 F~IFV1 [[.

To conclude this section, we summarize the results of it in the form of an algorithm, which computes the norm of the Hankel--operator with symbol V. Step 1:

Find

a

realization

[A,B,C,0]

of

the

strictly

proper

and

anti-stable

p~

of V(s): V(s) = [A,B,C,0] + (a matrix in M(IP,H**)). (In our particular case this factorization is given by T + XY)C T L -1/2,H-1/2B T ,0] + (H-1/2DTL1/2) V(s) = [-A¢,(I in accordance with formula (5.4).) Steo 2:

Solve the Lyapunov equations (5.42) and (5.43) for L c and L o.

Step 3:

Then II F v II eigenvalue of LcL o.

=

~/9%~x(LcLo),

where

Xmax('LeLo)

is

the

largest

To calculate the maximal stability radius, we only have to take one extra step:

Step 4: The

w~l = (1 + l[ F v II2 ) -1/2 -- (1 + X.~(L~Lo) ) -1/2. algorithm

described

above

(without

step

4),

(5.46) is

implemented

in

MATLAB-function hankel. For the details of this implementation we refer to appendix C.

the

65

5.4. Solving the Nehari-pmblem In the last section we derived a method to compute w~1, the maximal stability radius that can be achieved by one compensator. To do so, we had to calculate the value cc of the infimum Cc

i nf

=

R ~ ~q~'n(s)

II V+R II-- II r V II.

But we are not only interested in the stability radius itself, but also in the compensator C that realizes this bound. From corollary 4.2.4. we know that for the construction of such a compensator, we need a matrix R such that II V+R II -- II r y

II. Nehari's theorem (theorem

5.3.1.) assures that such a matrix in fact exists. The problem to find such an R is called a

Nehari---problem. In

this

section

we

will

describe

a

method

to

solve

the

Nehari--problem

in

the

sub--optimal case. I.e. for each e > 0, we give a construction of a matrix R e e Rttm***n(s) such that

(5.47)

II V+R e II ~ II r v II (l÷e).

However, it is also possible to construct an exact solution to the Nehari-problem. In section 6.2. we will return to this subject. At first we replace the plus-sign in the infimum by a minus-sign. I.e. for each E > 0 we search for a matrix R e e M(P,I-I**) such that II V - R e II ~ I1 r y

II (l÷e). In this way we

get a standard Nehari-problem as described in most literature. So in the end we have to replace R e by - R e to solve the original problem. Analogous as in the last section it is possible to decompose V in a stable and a strictly proper anti-stable part: V(s) = Vl(S) + V2(s) = [A,B,C,0] + (a matrix in M(NI-I**)) with A anti-stable. The matrix V2(s ) in M(P,I-I,,) can be approximated exactly in M(P,I-Io,) (by itself),

so in

this

way

we can reduce

the sub-optimal

Nehari-problem

to the following

problem: "Let V 1 be completely anti---stable, i.e. V 1 is strictly proper and V~ e M(P,I-I ). Find for each e > 0 a matrix R e a M(P2-I**) such that II V1-R e II -< II rv1

II (l+E).- The original

problem can then be solved by adding the matrix V 2 to our solution R e. We will now present a solution to the Nehari-problem as stated in [4, sec.8.3.]. A proof of it would take us too far afield. We therefore omit it here, and refer for it to [4].

66

Let

II rvx II

Vl(s)

[A,B,C,0]

=

with

A

anti-stable

(so

V~

e

M(~I**)),

and

suppose

< 1. L e t L c and L o be the unique solutions to the Lyapunov equations

AL e + LeAT -- BB T

(5.48)

ATLo + LoA = cTc.

(5.49)

Define N := (I-LoLe) -1, and the matrix L by

L=[LL21

(5.50)

t3L4

with

L~(s) = [A,-L,.NCT,C,rl

(5.51a)

La(s) = [A,NTB,c,0I

(5.51b)

L3(s ) = [-ATINcT,-BT,0]

(5.51c)

L4(s ) = [-AT,NLoB,BT,I].

(5.51d)

Then we have the following theorem, which shows that there is not a unique solution to the Nehari---problem, but a whole set of solutions. THEOREM

5.4.1.

Let

V~ e M(RH**). The

II Va-R II ~

set

V1 • of

M(RL**), strictly proper, and all matrices

e

MtqRH**) such

that

I[ FV1 V1-R

e

{1 <

1, and

M(~H,)

and

i is given by:

R = V 1 - XIX~I, with

R

suppose

(5.52)

[ Xl

where Y e M(RH**),

II Y I1,, -< 1.

With help of this theorem we can now solve the Nehari--problem in the sub---optimal case. Let V i • M(RL**), strictly proper and such that V~ e M(RH**); so V 1 is anti-stable.

Suppose II FV~ II

= ~. Let e > 0. Then the norm of the Hankel--operator with symbol W 1,

defined by

W I := ~

1

V1

(5.54)

67 is smaller than one:

II r w a II = II r

1

v~

11 = ~

1

1

II rv~ II = l ~

< 1.

So we can apply theorem 5.4.1. to W r This gives us a parametrization of all matrices R e M(RI-I**) such that

II W:R Now

suppose

II ~ 1.

R x is

a

matrix

that

belongs

to

that

set,

so

1] W 1 - R t

11 -< 1,

with

R 1 e M(RI-I**). Then we also have:

II va - a(l+e)R1 II = II Ct(l+E)Wl -- CX(I+e)Rt II = = ct(l÷e) II W1-R, II ~ cx(a+e) - II ryl II (l+e).

(5.55)

So a ( l + a ) R z is a sub--optimal solution to our Nehari-problem. construct

the

set of

sub--optimal

solutions

to the

It is therefore possible to

Nehari--problem

for the

matrix

V 1 by

multiplying the set of solutions for W 1 by the factor ct(l+e). With help of (5.50) we can give a formula for the solution R in theorem 5.4.1. in terms of V 1 and L1,L2,I.,3 and L,t: X1

.,_

=

= [ L1Y+/'~

L2

L3Y+L 4 ]" So Y • M(~J-I**), [I Y ]] < 1.

R = V 1 -- (LIY+L2)(I.,3Y+Ln) -t,

(5.56)

Now we are often not interested in the whole set of solutions, but only in a particular one. Formula (5.56) suggests that the simplest solution to be calculated is given by the choice of Y = 0. Then (5.56) becomes R = V 1 - L2L~'t. We

(5.57)

now summarize the results of this section in the form of an algorithm which

calculates a sub--optimal solution to the Nehari--problem, with help of formula (5.57). Step 0:

Give a matrix V e M(RL,), and a tolerance level ~ > 0.

SteP 1:

Find

a

realization

[A,B,C,O]

of

the

strictly

proper

and

anti-stable

part

68 of V(s): V(s) = V1(s) + V2(s)

where

Vl(S)

=

with

[A,B,C,0]

A

anti-stable,

and

V2(s )

a

matrix

M(P,H**). (In our particular case we have: VI(s) = [_AT,(I+XY)CTL-I/2,H-1/2BT,0]

V2(s) = (_H-1/2DTLI/2)).

II FV1 II by

Step 2:

Calculate ct :=

the method of section 5.3.

Step ~:

Set W 1 := ~

Step 4:

Solve the Lyapunov equations

1

V 1, so Wl(s) = [A',B',C',0], with A' = A, B' = B, C' = ~ - 1~ C.

A'L c + LeA 'T = B'B 'T A'TLo + LoA' = c ' T c '. Step 5:

Set N := (I-LoL~)q and define

l..2(s) := [A',NTB',C',0] L4(s) := [-A'T,NLoB',B'T,I]. Stet) 6:

A sub--optimal solution R e to the Nehari-problem is then given by Re(s )

-- V2(s) + ot(l+E)(Wl(s)-L2(s)L4(s)-I) = = V(s) - cKl+e)l_~(s)L4(s)-a.

(5.58)

Remark that in the algorithm described above two pairs of Lyapunov equations have to be solved, which are highly related. For the computation of

II rvl II

we have to solve the

Lyapunov equations with the realization [A,B,C,0] of V r Let L e and L o denote the solutions to these equations:

ALc + LeA T = BB T

(5.59)

69 ATLo + LoA = c T c .

(5.60)

Later on, we have to solve the Lyapunov equations for a realization [A',B',C,0] of W 1. Suppose L'e and L 'o denote these solutions, then we have

A'L~ + L~A 'T -- B'B 'T

(5.611

A'TLo + LoA' = c ' T c .

(5.62)

Now we know that the realizations [A,B,C,0] of V 1 and [A',B',C',0] of W, are connected by the following relations:

A' = A;

B' = B;

C' =

1 ~--0-~ C.

(5.63)

Substitution of (5.63/ in (5.61) and (5.62) gives

A L 'c + L~A T = BB T

ATLo + LoA = ( a - ~ )

(5.641 2 cTc.

(5.65)

Combination of (5.64) with (5.59), and (5.65) with (5.60) immediately yields L'¢ = L e (5.66) L; = (a-~)

2 L o.

(5.67)

So we conclude that we only need to solve the Lyapunov equations once, to solve the Nehari-probIem in the sub--optimal ease. Putting together the algorithm from section 5.3. (for the norm of a Hankel---operator) and the one described above, and using the relations (5.66) and (5.67) between the solutions of the Lyapunov equations, we derive the following algorithm. .~tep 0:

Give a matrix V E M(RL,), and a tolerance level ~ > 0.

Stet~ h

Find a realization of the strictly proper and anti-stable part of V(s): V(s) -- Vl(S) + V2(s), where

M~J.

Vl(s )

=

[A,B,C,0]

with

A

anti-stable,

and

V2(s)

a

matrix

in

70 StCD 2:

Solve the Lyapunov equations

ALe + LeAT = BB T ATLo + LoA = c T c . Step 3:

Calculate the norm of the Hankel---operator with symbol V:

,~ := II r v II = ,/~.~,0~Lo~. Step 4:

SteD 5:

Set N := fl - ( _ _ ~ ) 2

LoLe)-t, and define:

L2(s) := [ A , N T B , a - - ( ~ C,Ol

(5.68)

L4(s) := [-AT,(~-(~-~)2NLoB,BT,I].

(5.69)

A sub-optimal solution R E to the Nehafi-problem is then given by Re(s) = V(s) - ~(l+e) L2(s)L4(s)-I.

(5.70)

This method is used in the MATLAB-function hanneh to solve the Nehari-problem. Given a state--space realization of a matrix V e M(~L**), and a tolerance level e > 0, hanneh computes the norm of the Hankel---operator with symbol V, and a sub-optimal solution to the corresponding Nehari--problem, within the desired accuracy. In the computer implementation there is still one improvement; it uses the formula R~(s) = V(s) -

[A,NTB,c,o]

* L4(s)-:

(5.71)

instead of (5.70). This formula however, is immediately clear from (5.68), (5.70) and the fact that

(x(l+E) Lz(s) -- (x(l+E) [A,NTB,~--~-~ C,0] = [A,NTB,c,0]. A state-space realization of Re is then calculated with help of the addition, multiplication and inversion formulae of page ix. For more information about the MATLAB-function hanneh, we refer to appendix C.

71 Finally we recall once more the la'ivial fact that for the solution of the original problem (find an R ~ M01I--/**) such that II V+R II ~ the matrix R we have found with this algorithm by - R .

II r v II (1+0 1, we have to replace

5.5. An algorithm to compute an optimally robust controller and its stability radius In the last three sections we have developed all the tools we need for the construction of an algorithm that solves the problem of optimally robust control in the dynamical case. In this section we put all these pieces together to derive such an algorithm. Suppose P G ~*m(s), and let [PA,PB,PC,PD] be a minimal realization of P. Assume PA

¢

[ ].

With

the

method

described

in

theorem

4.1.1.

and

implemented

in

the

MATLAB-function ncoprfac, we can calculate state-space realizations of an n.r.b.f. (D0,N0) and an n.l.b.f. (~o,~o) of P and of matrices Y0,Z0,~?0 and ~0 such that (4.7) holds. With the MATLAB-function speccopr we are also able to compute a state-space realization of the matrix V = D~7_~+N~ 0 by the method of theorem 5.2.1. Given a state--space realization of the matrix V and a tolerance level e > 0, the MATLAB-function hanneh computes a = ]1 F V }1 =

and

a

II V-R~ t l -

state-space

i nf II V - R 11 = i nf ]l V+R II R ~ RH~*n(s) R e ~Im'n(s)

realization

of

a

matrix

Re

~

(5.72)

~J-Im*n(s)

such

that

II r y II (1+~). The used method is explained in the last section. Now

combination of (5.72) and (5.46) immediately yields the following formula for the maximal stability radius w~k

w~X = (1 + II r v 112)-t/2 -- (I ÷ az)-a/L

(5.73/

To realize this bound we have to construct a state-space realization of a matrix R m-Pjn(s) such that II V+R II = I I r v II, and substitute this solution to the Nehari---problem in formula (4.46) or (4.56) to derive an optimally robust contxoller. The MATLAB-function hanneh however, delivers us a state-space realization of a sub--optimal solution R e to the Nehari-problem, i.e. a matrix R e ~ RHm***n(s) such that t[ V - R e II-< Substituting - R e into (4.56), we find a sub--optimally robust controller Ce:

CE

=

(~,0-DoR£)(~0+NoR£)-q.

II r y H (l+e).

(5.74)

72 At this moment state--space realizations of all the matrices at the right-hand side of (5.74) are available, so it is possible to calculate a state-space realization of C e with the addition, multiplication and inversion formulae of page ix. At this stage we remark that the state-space realizations of the matrices lq0,/30,Y0 and Zo have not been used in the computation of C e. Since the matrix V is calculated directly, using the method of section 5.2., these matrices are not explicitly needed any more. With this knowledge the MATLAB-funetion speccopr is developed. Given a state-space realization of a plant P, it only returns the state-space realizations of the matrices N0,D0,Y0 and 7_~ with the properties described in theorem 4.1.1., and a state-space realization of V. For more details about the MATLAB--function speccopr we refer to appendix B. We finally remark here that the function speccopr makes the original MATLAB-function ncoprfac superfluous for the algorithm we are now developing. There is now only one problem left. Because we didn't calculate an exact solution to the Nehari---problem, but a sub--optimal one, we can't expect the controller C e to realize the maximal stability radius w~1 exactly. So to check whether the controller C e is indeed sub--optimally robust in the sense that it realizes w~1 within an accuracy level of e, we have to calculate the exact stability radius w-1 belonging to C e. This can be done with formula (4.58). From theorem 4.3.3. we know that for all R • RHm*"(s):

w =II

I ~'°-N°R 1

II = ~/1 + II V+R 112.

~o+DoR

So in our case (with R = -Re) we have:

w = ~/1 + II V-R~ Now, II V-R e II2 ~ II r v

112.

II2 (1÷E)2,

w = J l + I[ V-R~

(5.75) so

112 _< J l + II rv

../(1+e)2 + II r v

II2 (1+8) 2

-<

It2 (1+~)2

(l+e)J1

II2 = (5.76)

= (I+E) we. And because ~ 1

+ II r v

> l--e, this immediately yields

w-a > ~

1

-_ ~

I

w~l > w~l (l--e).

(5.77)

73 So Ce is indeed a sub-optimal solution to the problem of optimally robust control. Again we summarize the results of this section in the form of an algorithm.

Et.¢~ O:

Give

a

minimal

realization

[PA,PB,PC,PD]

of

a

plant

P

e

~'r~(s),

and a tolerance level e > 0.

Step ! :

Calculate state--space realizations of the matrices defined in theorem 4.1.1. and of V, defined in (5.4).

Step 2:

Compute

II

Fv

II

and

a

state-space

N0,D0,~ 0

realization

of

and

a

~o

as

matrix

Re E ~Im**°n(s) such that 11 V-R e II < [I F V [I (I+E) by the method of section 5.4.

Step 3:

Set wit := (1 + II r v

Step 4:

Compute

a

ll2)-1/2.

state-space

realization

[CA,CB,CC,CD]

of

the

sub--optimally robust controller CE:

Ce = (~0-DoRe)(~0+NoRe)-i with the addition, multiplication and inversion formulae of page ix. The MATLAB-function robstab is the computer implementation of this algorithm. To take the last step, it uses the MATLAB-funcdons ffadd, tfmuh and tfinv, which are implementations of the addition, multiplication and inversion formulae of page ix respectively. More information about these algorithms can be found in appendix E. For the details of the MATLAB-funetion robstab, we refer to appendix C. For the sake of completeness we here recall the fact that the Nehari--problem we have to solve in our algorithm doesn't have a unique sub-optimal solution, but a whole set of sub--optimal solutions (see (5.56)). Each one of them gives rise to an other sub--optimally robust controller. For each e > 0 it is even possible to parametrize the set of all sub--optimal controllers which realize the maximal stability radius wit within a tolerance level of E. The algorithm we developed however, calculates only one particular compensator. Although the algorithm of this section delivers a sub-optimal compensator to the

74 problem of optimally robust control, it still solves the problem in a certain sense. We can see this as follows. Given a plant P, we can calculate the maximal achievable stability radius w-~l; w~~ only depends upon P. And when a tolerance level e > 0 is given, we have seen that we can construct a compensator C E which realizes a stability radius w-1 such that (l--e) w~1 ~ w-1 ~ w~L

(5.78)

So the maximal stability radius w~l can be approximated to any wished accuracy. In practice this means the following. Suppose the plant Pk is a perturbation of P, and 5(P,P~.) < w~l. Then we know there exists a compensator C that stabilizes both P and Pg. But with our algorithm we can also construct such a controller C. Since 8(P,P~) < w~l, there exists an e > 0 such that 5(P,Pz) < w~l(1-e). Choose such an e > 0. Then the controller C e, computed by our algorithm, stabilizes not only P but also P~. because 5(P,P~) < w~l(1--e) -< w-l. In this way we can say that our algorithm solves the problem of optimally robust control. Finally we remark that the performance of this algorithm will be discussed in section 6.4. This will then be illustrated by a little design example.

6. REDUCTION OF THE ORDER OF THE COMPENSATOR In the last chapter we developed an algorithm for the solution of the problem of optimally robust control. Unfortunately the compensator, calculated by this algorithm, is of a very high order. This is the main reason why this solution is hardly applicable in practice. In this chapter we try to overcome this problem. We develop two new algorithms, of which the last one turns out to reduce the order of the optimally robust controller drastically. In

the

first

section

of

this

chapter

we

determine

the

order

of

the

compensator,

calculated by our original algorithm. In this way we try to get some insight in the reasons for the inflation of the order of the conla'oller. With this knowledge we derive an algorithm for the computation of an optimally robust controller of a much lower order in the second section. In section 3 we introduce a fully different method which enables us to decrease the order of the solution even much more. Finally, in section 4, we present

a little design

example to compare the performances of the different algorithms. 6.1. The order of the compensator Suppose

P

E M(~p(s))

and

let

[PA,PB,PC,PD]

be

a

minimal

realization

of

the

transfermatrix P. Assume PA # [ ]. The order of P is then defined as the size of the square-matrix PA, the system-matrix of the plant. Assume that PA is an nxn matrix. Then we can determine the order of the compensator C, calculated by our algorithm, in terms of n.

First we have to compute a state-space realization of the matrix V, by the method described in theorem 5.2.1. So the system-matrix of V is - A T, with A e -- (PA)--(PB)F, and F defined

as in

theorem

5.2.1.

Clearly A~ is also

an

nxn

matrix,

so this

state-space

realization of V has order n. For this matrix V ~ MtqP,L,,) we then have to solve the Nehari-problem. From formula (5.70) we know that a solution with an accuracy level of 8 is given by Re(s) = V(s) - a(l+e)L2(s)L4(s)q

(6.1)

with 1---~ L2(s) = [A,NTB,~

C,OI

L4(s ) = [-AT,(~--(~)2NLoB,BT,I] where A is the system-matrix of the anti--stable part of V. Since in our case this matrix is T - A e, both L 2 and L 4 have order n. From the multiplication formula of page

ix, it is

immediately clear that the order of the product of two transfermatrices, multiplicated by this method, is the sum of the orders of the single ones. For the addition of two transfermatrices

76 we also have this property: the order of the sum is the sum of the orders. However, when we invert a transfermatrix by the formula of page ix, the order doesn't change. When we use this knowledge to calculate the order of R e, the sub--optimal solution to the Nehari-problem (formula (6.1)), we get the following. L,l(s ) has order n, and so does La(s)'-l. Since L2(s ) also has order n, the product a(l+e)L2(s)L4(s)-I has order 2n. So when we add V(s), of order n, to --~(l+e)L2(s)L4(s)-l, the sum Re(s) has order 3n. So we see that the state--space realization of the solution R e to the Nehari---problem, calculated by the MATLAB-function hanneh, is of the order 3n. This solution R e to the Nehari--problem is then substituted in formula (5.74) to get a sub-optimally robust conl~oller Ce:

(6.2)

C e = (7.0-DoRe)(~0+NoRe)q

where D0,N0,Z 0 and ~0 are defined as in theorem 4.1.1. By formula (4.6) we know that Ae is the

system--matrix

of all these

transfermatrices,

where

A e = (PA)--(PB)F,

so all these

transfermatrices have order n. Using the order properties of the addition, multiplication and inversion formulae of page ix, described above, we get: the order of (~0+NoRe)q is equal to the

order

of

(~0+NoRE),

and

this

is

n+(n+3n) = 5n.

The

order

of

(~,0-DoRe)

is

also

n+(n+3n) = 5n. So the order of C e, the product of these transfermatrices, is equal to the sum of their orders: 10n. So the order of the state-space realization of the sub-optimally robust controller calculated by the MATLAB-function robstab, is ten times as high as the order of the original system P. From the exposition above, we can see that there are two reasons for the inflation of the

order

of

Nehari--problem.

the

compensator.

At

first

the

order

is

increased

when

we

solve

the

Given a strictly proper V e M(IRL**), such that V* e M(RH**) (so V is

anti-stable) and of order n, our algorithm delivers a solution to the Nehari-problem of order 3n. Secondly the order is increased because we have to substitute this solution in formula (6.2) and use the formulae of page ix, which increase the order very rapidly.

We even

remark that this is also the main reason why our solution to the Nehari-problem is already of such a high order. Finally we remark here that the situation would even be worse when we didn't have formula (5.4) or (5.5) for the construction of a state-space realization of the matrix V. With this method we find a state-space realization of the order n. When the realization of the original plant P is minimal, we

even know

(theorem

5.2.2.)

that this realization is also

minimal. Therefore an n th order realization of V is the best we can get. When we calculate V by the original formula V = D ~ 7 - 0 - N ~ 0, where all the transfermatrices on the fight-hand side have order n, the order of the state-space realization of V is (n+n)+(n+n) = 4n. And with this realization of V we have to solve the Nehari----problem.

The situation is then even

much worse. So theorem 5.2.1. already reduces the order of the solution drastically.

77 6.2. The method of Glover To decrease the order of our compensator we develop in this section an other method for the solution of the Nehari-problem. In the last section we saw that the method described in paragraph

5.4.

was

one

of the

main

reasons

for the

inflation

of the

order of the

compensator. The method of this section is based upon an article of Glover [7], and delivers an exact solution to the Nehari--problem. When the order of the anti--stable matrix V e M(P,L**), which

has

to

be

approximated

by

a

matrix

R e M(P,I-I**), has

order

n,

the

state--space realization of the solution, constructed by this method is of order n-1. In this way it is possible to reduce the order of the state-space realization of the optimally robust controller to 6n-2. Before we concepts,

which

describe will

be

the

method

needed

later

of Glover, we in

this

first have to introduce

section.

We

start

with

the

some new

definition

of

Hankel-singular values. DEF.6.2.1. Let V e M(RL**), and such that V* •

M(PSI**). Let [A,B,C,D] be a minimal

realization of V. Then A is anti--stable. Let L c and L o be the unique

solutions to the

Lyapunov equations

AL e + LeA T

=

BB T

ATLo + LoA = c T c . Suppose the eigenvalues of LcL o are arranged in decreasing order, and let X~LcLo) denote the i th eigenvalue of L~L o. Then the Hankel-singular values of V(s) are defined as

oW(s)) := /-~U=Lo).

(6.3)

So by convention at.pC(s)) > a~.l(V(s)). Remark that this definition makes sense because from theorem 5.3.2. and 5.3.3. we know

that

all

the

eigenvalues of LcL o are real

and

nonnegative.

When

the

realization

[A,B,C,D] of V is minimal one can even prove that all the Hankel-singular values are larger than zero. With

this

knowledge

we

can

now

introduce

the

following

result

about

balanced

realizations. THEOREM 6.2.1. Let V e M(IRL**), and such that V* e M(P,H**). Suppose [A,B,C,D] is a

minimal

realization

¢)1 = a2 . . . .

° r > ¢~r+l > c~-2 >"> an > 0

of

V,

where be

the

A

is

Hankel--singular

anti-stable. values

of

V.

Let Denote

78 Z 1 = diag(t~r+l,Ow2,..,O'n) and define the matrix Z by

Then there exists a basis transformation T, such that for the minimal realization of V defined by V = [A',B',C',D'] := [T-tAT,T-1B,CT,D]

(6.5)

the following statement holds: The matrix Z is the solution to both the Lyapunov equations with respect to the realization (A',B',C'), i.e.

A'E + EA 'T = B'B 'T

(6.6)

A ' T z + EA' = c ' T c '.

(6.7)

A realization [A',B',C',D'] with this property is called a balanced realization of V. PROOF The proof is not very difficult and we refer for it to [7, see.4]. We remark here that the balanced realizations are a quite important tooL With this theorem

it

is

possible

to

find

a

realization

of

the

transfermatrix

V,

for

which

the

controllability and observability gramians have a very simple structure. We are now ready to state the main result of this chapter. It is an alternative version of a result in [7]. THEOREM 6.2.2. Let V e M(RL ), and assume V* ~ M(RH®). Suppose [A,B,C,D], with A anti-stable, is a minimal and balanced realization of V such that

AE + ~a~T = BB T ATE + ZA = c T c ,

with Y. =

1 CYlI

as defined in (6.4).

Partition A,B and C conformally with Y. as B1 ],

c = (c~,c2),

(6.8)

79 and define

X := F-I(o~AT, + ~;,An~,- o,cTuBT)

(6.9a)

:= r-,(r.,B, - o~cTu)

~2 :=

C1Z1

(6.9b)

otUBT

-

(6.9c)

A

D := D - o{U

(6.9d)

where U is a unitary matrix satisfying

B2 = c T u

(6.10)

and F is defined by

F := Z~ - c ] I .

(6.11)

Then the matrix R(s) e M(IRH ), defined by A

A

A

A

R(s) := [A,B,C,DI

(6.12)

is an exact solution to the Nehari--problem for the matrix V, i.e. we have

It V

-R

(6.13)

II = It FV II.

PROOF For the proof of this theorem we refer to [7, see.6]. The theorem stated above is a combination

of the theorems

6.1.

and

6.3.

in the

article of Glover. There is only one

difference. In [7] one wants to approximate a system V e M(~tt**) by an anti--stable one. We want to do it just the other way round: approximate an anti-stable system by a stable one. However, both cases are in essence the same and can be transformed into each other by replacing

V(s)

by

V(-s)

and

R(s)

by

R(-s).

In

our

version

of

theorem

6.2.2.

this

transformation has already been carried out. o Combining the theorems 5.2.2., 6.2.1. and 6.2.2. we develop a method to solve the Nehari---problem for the matrix V(s), defined in (4.57) exactly. Given a minimal realization of a plant P, we know by the theorems 5.2.1. and 5.2.2. how we can construct a minimal realization of the matrix V, belonging to P. The system--matrix of this realization is clearly anti-stable. But then, according to theorem 6.2.1., there exists also a balanced realization of V. And after balancing, application of theorem 6.2.2. yields a construction of a state-space realization of a matrix Rg(s) which is an exact solution to our Nehari-problem:

80

IIV-R

s II ~- l i t V II. A

The order of this solution is equal to n-r, the size of the system-matrix A of Rs(s), where r is the multiplicity of the largest Hankel---singular value of V. So in general the order will be n-1. This method for the exact solution of the Nehari--problem has been implemented in the MATLAB-function hannehgl. It makes use of the MATLAB-function ohkapp,

which is a

standard function in the MATLAB Robust Control Toolbox. This last function solves the problem as stated by Glover (so not our transformed version) conform Glover's method, but without the use of balanced realizations. The method used is given in [14] and bypasses the numerically often ill--conditioned balanced realization step, required in the original method. For the details of the implementation of the function ohkapp we refer to [2]. Now the MATLAB-function hannehgl first transforms our original problem to the one as stated by Glover, then uses the function ohkapp, and finally transforms the output of ohkapp back, to get an exact solution Rg to our original Nehari-problem. The norm of the I--lankel-operator with symbol V, which is equal to the largest Hankel---singular value belonging to V, is also computed. In appendix D more information about the MATLAB-function hannehgl can be found. Analogous to section 5.5. we now develop an algorithm for the computation of an optimally robust controller. This algorithm is completely the same as the one described in paragraph 5.5., apart from one difference: the Nehari-problem is exactly solved with the method of Glover instead of the sub-optimal solution method of section 5.4. So the exact solution R s to the Nehari-problem can be substituted as -Rg in formula (4.56) to calculate an optimally robust controller:

(6.14)

Cg = (~-.o-DoR~)(~0+NoR~)-k

As a consequence the maximal stability radius wit is really achieved by this controller. So the method of Glover delivers a true solution to the problem of optimally robust control. In the resulting computer algorithm,

implemented in the MATLAB-function robstgl,

there are two differences with the one developed in section 5.5. (the MATLAB-function robstab). First of all no accuracy level e > 0 is needed, because the Nehari--problem is solved exactly. Secondly the MATLAB-function hanneh is replaced by the MATLAB-function hannehgl. For the details of the function robstgl, we refer to appendix D. To conclude this section, we calculate the order of the state-space realization of the optimally robust controller, constructed by the MATLAB-function robstgl. Suppose

P

M(IRp(s)) and let [PA,PB,PC,PD] be a minimal realization of P. Let PA an n×n matrix, so P has order n. Analogous as in section 6.1. the state-space realization of the matrix V, related to P, and calculated with formula (5.4) is also of order n. For this matrix we solve the

81

Nehari-problem with Glover's method. In the general case this gives a solution Rg of the order n-1.

Substituting R s into (6.14) we find the following. (~0+NoRg)-I is of the same

order as (~o+NoRg) and since ~0 and N O are of the order n, this term is of the order n+(n+(n-l))

=

3n-1.

Completely

analogous

we

see

that

the

state--space

realization

of

(if.0-DoRa) is of the order n+(n+(n-1)) = 3n-1. So the order of the product is (3n-1)+(3n-l) = 6n-2.

So with the method of Glover we can construct a state--space realization of an

optimaIly robust controller of the order 6n-2. section 6.1.

we

draw

the

conclusion that

When

we

we compare this with the result of

have

decreased

different

method

the order of the

solution

considerably.

6.3. The method of Glover and McFarlanc In

this

section

we

describe

a

fully

for

the

computation

of

a

sub--optimally robust controller. It is based upon a recent article [8] of Glover and McFarlane. Given a plant P of order n, they give a method to construct a state-space realization of the order n of a compensator, which turns out to be a sub-optimal solution to our problem of optimally robust control. But before we can derive this result, we need some more theory. Therefore we start with the following lemma. (An alternative version of this lemma is given in [8], but without prooL)

LEMMA

6,3.1.

Let

P e

~*m(s),

and

suppose

[A,B,C,D]

is a minimal realization of P.

Assume A ~ [ ]. Let X, respectively Y, be the unique positive definite solutions to the Algebraic Riccati Equations (4..3) and (4.4). Define as in theorem 4.1.1.

A , := A - B F

with F := H-l(DTc+BTx)

A 0 := A - K C with K := ( B D T + y c T ) L -1

(where H = ( I + D T D ) ) (where L = (I+DD T) ).

Then the matrices (I+YX)-IY and (I+XY)-IX satisfy the following Lyapunov equations:

(I+YX')-IYAT + Ac(I+YX)-IY = _BH-IB T

(6.15)

(I+XY)-IXA o + AT(I+XY)-:X = --cTL-IC.

(6.16)

PROOF i) First we prove (6.15). Using the fact that L-ID = DH-1 (formula (5.12)) we derive from the definitions of A c and Ao:

ATc -- A To = ( A - B F ) T - (A-KC) T = c T K T _ FTB T _-

82 = c T L - t ( D B T + c y ) - (cTD+XB)H-IB T = = cTL-IDB T + c T L - I C y _ CTDH-IB T _ XBH-IB T = _- c T L - a C y - XBH-IB T.

(6.17)

Pre-multiplication of (6.17) with Y, addition of the term -BH-1B T on both right-

and

left-hand side, and a little rearranging yields y A T _ YA I - BH-1B T _ y c T L - 1 C y

=

-BH-1B T _ YXBH-1B T =

=. -.-(I+YX)BH-1BT.

(6.18)

Now from (4.13) we know that AoY = -yATo - y c T L - I C Y - BH-1B T.

(6.19)

Substituting (6.19) in (6.18), we get

YA T + AoY = -(I+YX)BH-IB T,

(6.20)

and pre-multiplicadon with (I+YX) -l yields:

(I+YX)-tyA T + (I+YX)-IAoY = -BH-1B T.

(6.21)

From (5.20) we know that (I+XY)A T = AT(I+XY), so Aa(I+YX) = (I+YX)A c and (I+YX)-IA o = Ae(I+YX)-'I. Substitution of this equality in (6.21) gives

CI+YX)-IyATe + Ac(I+YX)-IY = -BH-aB T, and this exactly formula (6.15). ii) The proof of (6.16) is almost completely analogous. When we transpose formula (6.17), change the roles of X and Y, and of BH-tB T and cTL-IC, and use (4.8) instead of (4.13), the proof of (6.16) follows immediately. The details are therefore omitted, a With help of this lemma it is possible to reduce the algorithm of section 5.3. for the computation of the maximal stability radius w~ 1 to a very simple formula. We already know that the maximal stability radius depends only upon the plant P. The next theorem describes how.

83 THEOREM

6.3.2. Let P e ~'m(s), and suppose [A,B,C,D] is a minimal realization of P.

Assume A # [ ]. Let X, respectively Y, be the unique positive definite solutions to the Algebraic Riccati Equations (4.3) and (4.4). Then the maximal stability radius wit, belonging to P, i.e. the radius of the largest ball around P that can be stabilized by one compensator C, is given by: 1

(6.22)

PROOF First we remark that in this proof we feel free to use the different matrix definitions given in the theorems 4.1.1. and 5.2.1. Now to prove (6.22) we simply carry out the algorithm for the computation of the maximal stability radius w~ 1 of section 5.3. Let P = [A,B,C,D] c R~*m(s). Then In'st we have to find a state-space realization of the strictly proper anti--stable part Vl(s ) of the matrix V(s), associated with P, as defined in (4.57). According to theorem 5.2.1. and because A c is a stable matrix, this is given by

VI(S) = t-A:,(I+XY)CTL-I/2,Hq/2BT0]. With this realization, we have to solve the Lyapunov equations (5.42) and (5.43). This yields:

- A : L c - LeA c = (I+XY)CTLqc(I+YX)

(6.23)

- A c L o -- L o A : = BH-1B T.

(6.24)

When we compare (6.24) with (6.15), we immediately see that the solution L o to (6.24) is equal to L o = (I+YX)qY.

(6.25)

To solve (6.23) we have to do some more work. First we pre--multiply (6.23) with -(I+X-y)-t and post-multiply it with (I+YX)q. This gives

(I+XY)qA:L~(I+YX)q + (I+XY)qL~A¢(I+YX)-I = - c T L - I C .

(6.26)

We now recall formula (5.20): (I+XY)A: = A:(I+XY). From this it follows immediately that ( I + X Y ) q A : -- A:fI+XY) q, and after transposition that Ac(I+YX)-I = (I+YX)-tA o. Substitution

84 of these formulae in (6.26) yields

AT(I+X'Y)qLe(I+YX) q + (I+X-Y)-ILc(I+YX)-IAo -- -cTL-1C.

(6.27)

Combining (6.27) with (6.16) we get (I+XY)-tLe(I+YX)q = (I+Xy)-IX.

(6.28)

Pre--multiplication of (6.28) with (I+XY) and post--multiplication with (I+YX) gives L e = X(I+YX).

(6.29)

With help of (6.25) and (6.29) we can now calculate LeLo: LcL o = X(I+YX)(I+YX)qY -- XY.

(6.30)

So we get II F V 11 = s/Z'max(LcLo) = ~#/--~-~('XY), and finally with formula (5.46):

w~l = (I + ([ F V

112)-x/u = (1 + ~.max(XY) )-I/2,

which is exactly formula (6.22). This completes the proof. [] Bcfore wc

introduce the result of Glover and McFarlanc

we

recall some

definitions

from chapter 4. In formula (4.45) we defined for e > 0 a neighborhood of a plant P0 e

~n*m(S) in the gap-topology: K(P0,e) = { P e ~n*m(s)

I

/5(P0,P) < ~ }.

(6.31)

However, it was also possible to describe a neighborhood of Po with help of perturbations of its n.r.b.f, (D0,No). According to (4.47) we defined for e > 0:

R(Po,e) = { P = 0N0+An)(D0+Ad) q ¢ lRn*m(s) I 11

An

I1<~ }, (6.32)

where (Do,No) is an n.r.b.f, of P0. Glover and McFarlane used this last description to tackle the problem of robust stabilization. They derived the following result.

85

THEOREM 6.3.3. Let P ~ ~*m(s) and suppose [A,B,C,D] is a minimal realization of P. Assume A # [ ]. Let X, respectively Y, be the unique positive definite solutions to the Algebraic Riccati Equations (4.3) and (4.4). Then the following statements hold: i) The largest radius r, such that R(P,r) can be stabilized by one compensator C is given by

r=~ = (1 + ~ ( X Y )

)-1/2.

(6.33)

1 (1+e). Define as before ii) Let e > 0 and define T := r,,-"-~

A c := A - B F with F := H-I(DTc+BTx)

(where H = ( I + D T D ) )

and (6.34)

W 1 :---- I + (XV-y2I).

Then the controller C, given by

C = [Ae-v/2(Wit)TycT(c+DF),'I~(wiDTycT,BTx,-DT],

(6.35)

1 stabilizes R(P,r=~x i~--~e). This controller C is of the s a m e o r d e r as the original plant P. PROOF For the proof we refer to [8, pp.825-826]. Now, when we compare formula (6.33) for r,,ax with (6.22) for w~l, we see that they are equal: r=ax = w~t.

(6.36)

So for both the characterizations of a neighborhood of P, the maximal stability radius is the same. This result is not very surprising when we recall the result of theorem ,I.2.3. There we proved that for 0 < E <_ 1, K(P,e) = R(P,e), so these neighborhoods are the same. But then 1) the controller that stabilizes R(P,r=, x +lr-g-E should also stabilize KfP,w~I

I

The validity of this last statement we can see as follows. Since C stabilizes P, C e S(P). Now let (130,1q0) be an n.r.b.f, of P and (Y0,Z0) matrices such that (4.38) holds. Then we know that there exists an R e M(RH**) such that

C = (Yo-R~o)q(Zo+R~ o)

(6.37)

86 1

(see formula (4.39)). Now C stabilizes R(P,w~ l ~+e), so according to theorem 4.2.1. we have

II ¢Y0-P~0),(z0+P~0) II ~ ws(l+~)-

(6.38)

But then formula (4.42) gives that the stability radius w--1 of C, in terms of the gap---topology is

w-a = II (Yo-RI~o),(Zo+RDo) II-1 m w~X ~

1

> w~a(x-~).

(6.39)

So C also stabilizes K(P,w~(1--~)) and we see that the controller C, defined by (6.35) is a sub---optimal solution to the problem of optimally robust control. The maximal stability radius w~l can be achieved by this controller to any wished accuracy by choosing E appropriately small. The great advantage of this controller, however, is its order: the order of the state--space realization of C given in (6.35) is equal to the order of the original plant P. With this conclusion the problem of order--reduction seems to be solved. robust

This method of Glover and McFarlane for the solution of the problem of optimally control has been implemented for computer purposes in the MATLAB-function

rbstglfar. Given a minimal realization of a plant P ~ ~*m(s) and a tolerance level e > 0, rbstglfar calculates the maximal stability radius w~1 belonging to P,

and a state-space

realization of a controller C that achieves this bound within the desired accuracy level, with help of the formulae (6.22) and (6.35). The order of this compensator is equal to the order of the given plant. Remark that for this computation only two Riccati-equations have to be solved. In contrast with the other methods no Lyapunov equations have to be solved (we solved them analytically in lemma 6.3.1.). Explicit state-space realizations of Bezout factorizations are also not needed any more. In this way the MATLAB-function rbstglfar is a very simple algorithm for the solution of the problem of optimally robust control. For implementation, we refer to appendix D. To

conclude

this

section

we

finally

remark

that

the

the

details

performance

of this of

the

MATLAB-function rbstglfar will be one of the subjects of the next section.

6.4. The performance of the algorithms In chapter 5 and the last two sections of this chapter, we developed three different algorithms for the solution of the problem of optimally robust control. In this section we wiU compare them to f'md out what the advantages and disadvantages of these methods are. This will be done with

help of

a little design example.

We choose a plant P e M0{p(S)) with transferfunction P(s) = s_--sT. 1 Then P is an

87 unstable plant (it has a pole in +1), and a

state-space

realization of P is given by

P(s) = [A,B,C,D] with A = B ffi C = 1 and D = 0. For this plant we computed the maximal stability radius w~l and a state-space realization of a (sub)--optimally robust controller C in five different ways:

i)

with the IVlATLAB-function robstab; tolerance level £ = 10s;

ii)

with the MATLAB-funetion robstab; tolerance level £ -- 10s; with the IvlATLAB-funetion robstgl;

iii) iv) v)

with the MATLAB-function rbstglfa~, tolerance level £ = 10-5; with the MATL/gB-function rbstglfar; tolerance level e = 10-8.

The results of these computations are listed in appendix F. Comparing these results, we immediately see that all the methods give the same value for the maximal achievable stability radius w~l; w]l = 0.3827. This is exactly what we expected. The three methods calculate w~l in almost the same way, they differ only in the way they compute a (sub)--optimally robust controller that realizes this bound. The example also illustrates the order inflation of the different methods. Our original plant P is of order 1. The compensators calculated by the MATLAB-function robstab are of the order 10, by robstgl of the order 4 (4=6-2), and by rbstglfar of order 1, in accordance with our predictions in the last sections. In the solutions for the sub--optimal ease (calculated by robstab and rbstglfar), we also encounter an other phenomenon: the controller turns out to be highly dependent upon the tolerance level £. When a tolerance level £ has to be achieved, the order of magnitude of the compensators is about 1/e. So, when the maximal stability radius has to be approximated very accurately, the compensators calculated this way have very large poles and zeros. This is very unsatisfactory because such controllers are hardly applicable in practice. The reason for this behavior is quite clear. Take for example the function robstab. In this algorithm we have to solve a Nehari-problem before we can compute a sub--optimally robust controller. For the computation of a sub-optimal solution to the Nehari-problem, we have to calculate (see the second algorithm of section 5.4,, step 4):

N = (I - ( _ ~ ) 2

LoLc).q,

(6.40)

where a = / ~ a x ( L e L o ) > 0. But then or2 is also an eigenvalue of LoL ¢. Suppose x is an eigenvector of LoL o corresponding to eigenvalue a 2. Then LeLox = a2x, and LoLeLox ffi a2Lox. So Lox is an eigenvector of LoL ¢ corresponding to the eigenvalue ot2, provided that Lox ~ 0. But this is quite clear from the fact that LeLox = a2x # 0. So a 2 is indeed an eigenvalue of LoL e.

88 1

Now define M := (I--(~--(~) ~ LoLo). Then N = M-k From the argument above it is clear that ( 1T~)2 is an eigenvalue of (~ 1----~) 2 LoL ¢, so 1--(I~lg) 2 is an eigenvalue of M. So, when ~ tends to zero, the matrix M becomes almost singular. For small tolerance levels e > 0, the matrix N is the inverse of an almost singular matrix. The computation of such an inverse is numerically unstable, and it is obvious that in this way the inverse of the tolerance level, e--t, gets into our solution. So the algorithm robstab is numerically not very satisfactory. When we look at the function rbstglfar, we see that almost the same thing happens. Here we have to compute the inverse of the matrix W t = I + (XY-72I), where

~t2 =

(l+)~max(XY))(l+e)2

(6.41) (see

formulae

(6.33)

and

(6.34)).

Let

x

¢

0

be an

eigenvector of XY corresponding to the largest eigenvalue ~'max of X-Y. Then we have: Wlx = (I + (XY-72I))x = x + Xmaxx -72x = = ((l+Xm~x) - (l+)~x)(l+e)2)x =

(6.42)

= (1+~,m~x)(-2e--e2)x.

So (l+)~max)(-2e-e2) is an eigenvalue of W], and when e tends to zero, this cigcnvaluc does too. For small tolerance levels E > 0, thc matrix W t becomes almost singular, and because we have to compute the inverse of W] to determine a sub-optimally robust controller C, we get into the same numerical problems

as before. From

the exposition

above it is quite

obvious that in this case the solution to the problem of optimally robust control depends upon the tolerance level e, in the way we have seen it in our little design example. Fortunately,

our

third

algorithm,

the

function

robstgl,

overcomes

these

numerical

difficulties. In this algorithm, the method of Glover is used to calculate an exact solution to the Nehari-problem and no tolerance level is used. So in this case we can't encounter the numerical problems of the other two algorithms, because they were caused by this tolerance level. When we look at the results delivered by this algorithm for our example, we see that the order of magnitude of the compensator is equal to the order of magnitude of the original plant. This result is very satisfactory. When we study the method of Glover more carefully we come to

the conclusion that

also other numerical difficulties are not very likely to

happen. The only problcm is the matrix F, defined in (6.11) by

r

which

has

to

= z~ -

a21,

be

inverted.

This

matrix

can

become

almost

singular

when

the

largest

Hankel-singular value is almost equal to the largest but one. This is not very likely to havven, and fully devendent upon the transfermatrix of the original plant P. We could say that in this case the problem is iU-condifioned.

89 Because of the numerical difficulties mentioned before, we can ask ourselves whether the compensators calculated by the MATLAB-functions robstab and rbstglfar still achieve the maximal stability radius w~1 within the desired accuracy. Because of rounding errors, the sub--optimality could be endangered. To cheek whether the calculated solutions are still (sub)--optimally robust,

we

simply

compute

the

stability

radius

w-1

belonging

to

the

plant--compensator pair (P,C) by the method given in theorem 3.2.5. First we calculate an n.r.b.f. (D0,N0) of P and an n.l.b.f. (lq0,150) of C. Since these factorizations are normalized, we have

II [ DNo o ] II = II ~0,1~0) II = 1, and the formula for the stability radius w-t as given in theorem 3.2.5. becomes

w-X = II (fSoDo + /~0No)-l t1-1

(6.43)

A state-space realization of the matrix U := 130D0+lq0N0 is easily obtained from the different algorithms we already developed. Normalized right- and left-Bezout factorizations can be constructed with help of the MATLAB-function ncoprfac. A state-space realization of the matrix U-t is then computed with the functions ffmult, ffadd, and tfinv, which are based upon the formulae of page ix. When

we

carry

out

this

procedure

for

the

compensator

calculated

with

the

MATLAB-function robstab with tolerance level ~ = 10--5, we get the following. U-1 is in this ease a transferfunction (P is a SISO-system)

and has a number of poles in the RHP.

Theoretically these poles have to be canceled out by zeros, and we see that this happens in our example. When we also carry out a rough pole-zero cancellation in the LHP, and then calculate the norm of U-t with help of the MATLAB-funetion hinfn (see [1]), we find

II u 4 II--2.6132, and w--1 = 1/[[ U -1 II = 0.3827. This is exacdy the value of w~1 we already calculated. So the compensator we have calculated is still sub--optimally robust. Using the same procedure for the compensator computed by the MATLAB-function rbstglfar (with tolerance level ¢ = 10-s) we observe almost the same behavior as before. The transfermatrix U-1 has, after a rough pole-zero cancellation , only very large poles and zeros. Its norm turns out to be equal to 2.6131. So again we find w-1 --- 0.3827. However, when we check the optimality of the solution calculated by the function mbstgl, we find immediately a transferfunetion U-1 with all its poles in the LHP. So it is possible to calculate the H**--norm without any pole-zero cancellation. The value of w-t turns out to be equal to 0.3827, as expected. Because in this ease no pole-zero cancellation is necessary, we conclude again that the MATLAB-funetion robstgl is numerically the most reliable one.

90

To conclude this section we finally summarize the results we have found. First of all the MATLAB-function

robstab turns out to be not very useful in practice. It gives solutions

of a very high order and the order of magnitude of the compensator dcpends upon the given tolerance level. The function robstgl, on the other hand, is numcricaIly very reliable, but the order of the compensator calculated by this algorithm is still rather high. This problem can be overcome

by the function rbstglfar, which calculates a controller of the same order as the

plant. This function however, has the disadvantage that it has the same undesirable numerical properties as the function robstab and produces solutions with very large poles and zcros. So we can give the following rule of the thumb for the use of the diffurcnt algorithms. When numerical

reliability is more

MATLAB-function

robstgl. W h e n

important

than

the

ordcr

of

the

compensator,

use

the

a lower order is preferred, use the function rbstglfar. The

disadvantage of a compensator with very large poles and zeros has then to be taken into account. So the choice of the method to use is a tradc-off between undesirable numerical properties and the order of the solution.

7. CONCLUSIONS In this final chapter we draw our conclusions. In the f'trst section we will summarize the results we have found so far. Later on, in the second paragraph, we give a short outline for eventual future developments in this field. 7.1. Snmmury In this paper we achieved the following: We described the problem of robust stabilization, and introduced the gap--topology to

-

measure distances between plants. We saw that the gap--topology

is compatible with the

problem of robust stabilization. We gave sufficient conditions for robust BIBO--stabilization in the gap--metric. -

- With help of the guaranteed bounds for robust stabilization we were able to find a solution to the problem of optimally robust control in the gap--topology. By maximizing the bound for robust stabilization, we found a (theoretical) method to calculate, for each plant P M(~(s)), the maximal stability radius belonging to it (i.e. the radius of the largest ball around P that can be stabilized by one controller C), and a controller C that realizes this bound. Most of the results, stated above, are due to Zhu (see [19]). In the derivation of them, right- and left-Bezout factorizations turned out to play a central role. These factorizations appeared to be a crucial tool in the development of the theory. In the second part of this paper, we made these results more concrete, and developed three algorithms for the computation of an optimally robust controller and the maximal stability radius for the case P ~ M(~p(s)). (In this case, P has a state--space realization P = [PA,PB,PC,PD].) Here we derived the following results:

-

Given a state-space realization of a plant P G M(~p(s)), we gave state-space

realizations of a normalized doubly-Bezout factorization of P, i.e. we constructed state-space realizations of a normalized right- and left-Bezout factorization of P ((Do,No) and (1~o,~3o) respectively) and of matrices Yo,Z0,~o and ~.~ such that

-Zo 15o

-

Yo ~o ]

I - Noo o

~o

I

0

,7..

We showed that the problem of optimally robust control can be reduced to a

92 Nehari---problem for a matrix V, related to the original plant P. Given a minimal state--space realization of P, we gave a method to conslruct a minimal state-space realization of this matrix V. -

With help of this matrix V, associated with P, we developed an algorithm that

computes

the maximal

stability radius belonging to P, and a state--space realization of a

controller C that realizes this bound. The controller C was calculated in two different ways: i)

with

help

method

of

a

sub--optimal

solution

to the

Nehari-problem,

found

with

the

of Francis (see chapter 5). This finally led to a sub--optimally robust

compensator of an order ten times as high as the order of the original plant P. ii)

by

the

method

of

Glover,

which

uses

an

exact

solution

to

the

related

Nehari-problem. This yields an optimally robust controller of an order almost six times as high as the one of the original plant. - Finally we gave a direct method to compute the maximal stability radius, belonging to a plant P, given a minimal state--space realization of it. We showed that the sub--optimaUy robust controller, based upon an other metric for plants, and given by Glover and McFarlane in [8], is also sub--optimally robust in our setting, i.e. in the gap-topology. The order of this compensator is equal to the order of the plant. -- We concluded our exposition with a comparison of the three algorithms. We saw that the method of Glover was numerically the most reliable one, but with the disadvantage of

a high

order

compensator.

The

method

of Glover and

McFarlane

did

overcome this

problem at the cost of undesirable numerical effects.

7.2. An outline for futm'e developments In section 6.4., where we compared the performances of the algorithms, we saw that they

all

have

their

own

advantages

and

disadvantages.

But

unfortunately

there

was

no

algorithm that combined all the advantages, yielding a compensator of the same (or lower) order as the plant in a numerically stable way, and without any drawbacks. Of course now the question arises whether such a compensator exists, and if so, how to calculate it. For a very special case, the first question can be answered affirmatively. THEOREM 7.2.1. Let P e 8p(s) a first order, strictly proper system with transferfunction

P(s) = s - ~3 cz'

with {3 ~ 0.

(7.2)

Then the maximal stability radius belonging to P is

wgl = ~/2(ot2 + 1~2 + ~

)

"

(7.3)

93

And an optimally robust controller of the order zero, that realizes this bound, is given by:

K = tx + , / " ~ +

~2

(7.4)

PROOF i) First we prove formula (7,3) for the maximal stability radius w~1. To do this, we simply apply theorem 6.3.2. A minimal state-space realization of P is given by P = [A,B,C,D] with A = ct, B = [3, C = 1 and D = 0. First we have to solve two Riccati equations (formulae (4.3) and (4.4)), which in this case are scalar: ctx + xot --~2x2 + 1 = 0 oty + yot -y2 + [~2 = 0.

(7.5) (7.6)

The positive (definite) solutions to (7.5) and (7.6) arc given by

x =

ct + 4°c2 + ~2 ; ~2

y = tz +

+ 9 2.

(7.7)

Now we know by formula (6.22) that w2 is given by

-- 1 + )~max(xY) = 1 + xy -- 1 + (a + ~

= 2 t e2 + ~ So, for the maximal

)2

+ e V ~ ' ~ - I ~ }.

stability radius w ~ I w e have

and this is exactly formula (7.3). ii) To prove that the controller K, defined in formula (7.4) is indeed optimally robust, we show that the stability radius w-1 belonging to the plant-compensator pair (P,K) is equal to w~1. To do so, we compute this radius w-t, by the method given in theorem 3.2.5. First we

construct an n.r.b.f. (D0,N0) of P. Since we

already solved the Riccati

equations, we only have to substitute the solutions of them in the state-space formulae of theorem 4.1.1. This yields

94 f = 9x = cc + ~/cc2 + ~2;

a c = a_gf = _ ~ +

92 .

9 (The capitals are replaced by lower ease letters, because they all represent scalars.) And using formula (4.6), we get:

D0(s) = 1 -

~f

s-ao

N°(s) = s-ae

With

help

=

= 1 -

a + ~

_

s + ~

s-

=

(7.8)

s+:--ar~

13

s +

(7.9)

of theorem

4.1.2.

it is possible

to

construct an

n.l.b.f.

(I~0,13o) of

the

constant transfermawix K. Application of formula (4.34) yields:

I5 o

= a+X.x.T) %

= (1 + ( a + ¢ - ~ ) ~ ) - V ~

= (7.10)

= .] 2(=2 + /32 + = ~ lq0

=

(I+KKT)-~I2K =

=~+:~

We

)

now

define,

191 ~/2(ct2 + 92 + ~ in

conformity

with

(7.11)

) (3.7),

Ao

:=

[Do] No

,

Bo

:=

(~o,1~o)

and

U o := BoA o. Because the Bezout factofizations are both normalized, we have

II Ao II = II [ No] Do

II= l,

II Bo II = II (15oNo) II-- 1. So the formula for w-t, the stability radius of the plant-compensator pair (P,K) becomes (see theorem 3.2.5.)

w--X = ( II ,% II II B0 II II U~t II )-1 = II U~I I1-1. Now we compute

Uo

U0:

= tioDo + l%No =

(7.12)

95

~/2(a2 +

/13213+I~

)

= ~/2(Ct2 + 132 + ~

)

+ (7.13)

w~l,

where the last equality obviously follows from formula (7.3) for w~t. With (7.13) we see immediately that Ust = wg, and so

w-a = II u~l 11-1 = w~a. So the stability radius w-t belonging to the plant--compensator pair (P,K) is equal to the maximal achievable stability radius w~l, and by consequence, K is an optimaUy robust controller for P. This completes the proof. D For the very special case of theorem 7.2.1. it is possible to calculate an optimally robust controller of a lower order than the original plant, in a very simple and numerically stable way. Remark however that the proof of the optimality of K was not constructive. In a certain sense, we introduced an apparently arbi~:ary compensator K, which turned out to be optimally robust. Of course, this is not the plain truth. We constructed K in such a way that the matrix U 0 (a scalar in this case) was constant, and this K turned out to be optimally robust. Naturally, now the question arises: can this result be generalized to higher order systems? The answer is ambiguous. When n is the order of the original plant, Glover and McFarlane have claimed that in theory there must exist an optimally robust controller of an order not greater than n - 1 (see [8, p.827]). The argument for this claim is the following. For the H**--control problem, which is highly related to our problem, almost the same question arose a few years ago. This question, however, has been solved in the mean time, and it turned out that we can construct a controller of a lower order than of the original plant in this case. So we expect this also to be possible for the robust stabilization problem. But if such a compensator really exists, how should it be calculated in a numerically stable way? And how can we construct a state--space realization of such a controller? These questions have not

been answered yet,

and further research

in this

area

seems to be

worthwile. In this way it should be possible to develop a numerically stable algorithm that computes an optimally robust controller of a lower order than of the original plant. The results of such an algorithm are then also very good applicable in practice.

APPENDIX A This appendix contains function-descriptions of the MATLAB-functions gap and dirgap. gap Purpose : Computation of the gap between two systems. ~ynopsi~ : [gl,g2,dist] = gap(A1,B1,C1,D1,A2,B2,C2,D2) Description : Given minimal realizations [AI,B1,CI,D1] and [A2,B2,C2,D2] of the plants PI and P2 respcctively, gap calculates gl, the directed gap between P1 and P2, and g2, the directed gap between P2 and P1, within an accuracy level, given by tol. The gap between P1 and P2, which equals the maximum of the two directed gaps, is returned in the variable dist. So in formulac we have:

gl = ~(PpP2), g2 = ~(P2,P1), dist = 8(PI,P2) = max { ~(PI,P2),~(P2,PI) }, where we didn't take the accuracy level tol into account. Algorithm : gap uses the method given in [6, oh.2] to calculate the gap between two systems. Let ( D i ,N i ) and ( ~ i ,[5i ) be normalized fight- and left-Bezout factorizations of Pi respectively (i = 1,2). Then, according to (2.17), we have: 8(PI,P2) = Q eM({]LH®) inf [] [ D1 Nl ] -

[ D2 N2

] Q IL.

(AA)

When we define G := D~D I + N~N t,

(A.2)

98 J1 := -I~2D1 + 132N1,

(A.3)

we can rewrite (A.1) as (see [6, p.5])

(A.4)

~(e~J'2) = Q ~i., MfPa-I )

Given state-space realizations of the plants Pt and P2 it is possible to construct state--space realizations of the matrices G and Jp as defined in (A.2) and (A.3) (with help of the functions norcor, ffmult and tfadd). Given these realizations of G and J1, the MATLAB-function dirgap is used to calculate the directed gap ~(PI,P2) within the desired level of accuracy. The calculation of the directed gap between P2 and Pl takes place in the same way, where G and Jl axe replaced by G* and J2 := -/qlD2 + 131N2 respectively, i.e.

]~(P2,P1) =

i nf

Q e MOP,H**)

II

[ G*j: Q ] ]].

(A.5)

Finally the value of dist, the gap between Pl and P2, is calculated as the maximum of the two directed gaps gl and g2. WARNING: gap is not suitable to compute gaps between large order systems, certainly not in combination with a small value for tol, the desired accuracy level. The computation becomes then very expensive. N.estcd fvnctions : norcor, see appendix B. dirgap; see appendix A. ffadd, tfmult; see appendix E. Proffram

."

function [gl,g2,dist] = gap(A1,B1,C1,D1,A2,B2,C2,D2,tol) Given minimal realizations of PI=[A1,BI.C1,D1] and P2=[A2,B2,C2,D21

gap calculates dist, the gap between

P1 and P2. where the tolerance for the accuracy is equal to tol. gl

and g2 are the th'rected gaps

between P1 and P2 respectively.

[DRA I,DRB I,DRCI,DRD I,NRA I,NRB I,NRCI,NRDI,.. DLAI,DLB I,DLCI,DLD 1,NLAI,NLB I,NLCI,NLD I]=norcor(AI,BI,CI,DI),

99

[DRA2,DRB2,DRC2,DRD2,NRA2,NRB2,NRC2,NRD2,.. DLA2,DLB2.DLC2,DLD2,NLA2.NLB2,NLC'2,NLD2] =norcor(A2,B2,C2,D2); [H 1,I-L2,H3,H4] ~mult(--DRA2',DRC2' ,-DRB 2',12RD2'.DRA1,DRB I ,DRC 1,DRD 1); [H5,H6,H7,I-IS]=tfmult(-NRA2',NRC2',-NRB2',NRD2'.NRA1,NRBI,NRC 1,NRD 1); [GA,GB,GC.GD]=tfadd(H1 ,H2,H3 ,H4,H5,H6,HT,H8 ); GSA----GA';GSB---GC';GSC=-GB';GSD--GD'; [H 1,H2,H3./-I4]=ffmult(NLA2,NLB2,-NLC2,-NLD2,DRA1,DRB 1,DRC 1,DRD 1); [HS,H6,H7,I-IS]=ffmult('DLA2,DLB2,DLC2,DLD2,NRA 1,NRB 1,NRC1,NRD1); [JA I,JB 1,JC1.JD 1]=ffaddfl-II,H2,H3.1-I4,1-15,H6,H7,HS); [H 1,H2,H3,H4]=tfmult(NLA1,NLB 1,-NLC1,-NLD 1,DRA2,DRB2,DRC2,DRD2); [H5,H6,H7 ,H8]---tfmultfDLA 1,DLB 1,DLC 1,DLD 1,NRA2,NRB2,NRC'2,NRD2); [JA2,JB2,JC'2,JD2]=ffadd(H 1.H2,H3,H4,H5,H6,H7,H8); gl=dirgap(GA,GB,GC,GD,JA1,JB 1,JC I,JD 1,tol); g2=dirgap(GSA,GS B,GSC,GSD,JA2,JB 2,JC2.JD2.tol); dist=max([gl g2]);

dirgap Plll'DOSe

Computation of the directed gap between two systems, dirgap is a special purpose algorithm, to be used in the MATLAB-function gap. Synopsis : alpha = dirgap(GA,GB,GC,GD,JA,JBjC,JD,tol) Descrit~tion : Given realizations [GA,GB,GC,GD] of a matrix G, and [JA.IB,JC,JD] of a matrix J. dirgap computes, under the assumption that inf 11 |r G ~ _ Q |a H < 1, L " J Q ~ M(RI-I**) the value of this infimum (within an accuracy level of tol), and stores it in the variable alpha:

100

alpha =

The

relation

of

i nf II [ G J Q ] l l Q e M(~,I-I**) this

formula

with

the

(A.6)

directed

gap

is

already

explained

in

the

function-description of the function gap. Note that in the case where G and J are defined as in (A.2) and (A.3), the value of alpha is equal to a directed gap, and so smaller or equal to one. So in this ease (the only one we are interested in), the assumption

inf Q e M(P,I-I**)

11 ra I G ; Q [ [ l ~ l , L

~

3

is always satisfied. Algorithm : dirgap is a computer-implementation of the algorithm for the calculation of formula (A.6) as given in [6, p.6], which is based upon the algorithm of Francis described in [4, sec.8.1.]. It proceeds as follows. From formula (A.6) and our assumption it follows immediately that

II z II ~ alpha Now define IX :=

II

J

II

<

1.

and suppose IX < "y < 1. Let F T be a spectral factor of the matrix

~2I -- J'J, i.e.

F 7 ¥ = 72I - J'J,

(A.7)

with F~, and F~I e M(RH**). In [4, see.7.3.] it is proven that such a factorization really exists. Now it can be shown that (see [4, see.8.1.]) that

i nf I[ [ G J Q ] i[ < .y Q ~ MO~-I~) where

~,

II I"GF.~I II -< 1.

FGF~I denotes the Hankel-operator with symbol GF@

calculate possible, accuracy out this

(A.8)

Because we know how to

the norm of this Hankel--operator (see the MATLAB-function hankel), it is now with help of formula (A.8), to approximate the value of alpha to any wished by performing an iterative search on the interval [ix, l]. Bisection is used to carry search.

101

WARNING: Remark that in each step of the algorithm a quite expensive spectyal factorization has to be carried out. Therefore this method is not very good applicable for matrices of a large order, and certainly not in combination with a small value for tol, the desired level of accuracy. Nested functions : hinfn; see [1]; hankel; see appendix C. tfinv, ffmult; see appendix E. ProLrl-am :

function alpha=dirgap(GA,GB,GC,GD,JA,JB,JC,ID,tol) Given realizations IGA,GB,GC,GD] of G and IJAjB,JC,ID1 of J. dirgap calculates:

''

Q e MdRH**)

I,[

within an accuracy of tol. (Tlu's under the assumption alpha < 1.)

[mu,freq,l]=hinfn(JA,JB,JC,JD, 1e--8,0); under=mu;upper=l; while ((upper--under)>tol) gamma=(upper+under)/2; [FA,FB,FCh,FDh]=sfl(JA,JB,((1/gamma)*JC),((1/gamma)*JD)); FC=gamma*FCh;FD=gamma*FDh;

[FIA,FIB,FIC,FID]=tfinv(FA,FB,FC,FD); [RA,RB,RC,RD]=tfmult (GA,GB,GC,GD,FIA,FIB,FIC,FID); han=hankel(RA,RB,RC,RD); if hart _< 1 upper=gamma; else under=gamma; end end alpha=(uppcr+under)/2;

APPENDIX B This appendix contains function--desefiptions speccopr.

of the MATLAB--functions neoprfac, norcor and

ncopl~ac:, n o l ~ Pul'p_ osg :

Normalized doubly-Bezout factorization. Synopsis : [DRA,DRB,DRC,DRD,NRA,NRB,NRC,NRD,,. YRA,YRB,YRC,YRD,ZRA,ZRB,ZRC,ZRD,.. DLA,DLB,DLC,DLD,NLA,NLB,NLC,NLD,.. YLA,YLB,YLC,YI_.D,ZLA,ZLB,ZLC,TJ.D] = neoprfac(A,B,C,D) [DRA,DRB,DRC,DRD,NRA,NR.B,NRC,NRD,.. DLA,DLB.DLC,DLD,NLA,NLB,NLC,NLD] = nomor(A,B,C,D) DescripIion : Given a minimal realization [A,B,C,D] of a transfermatrix P(s), ncoprfac computes state--space realizations

of a

normalized

fight-

and

left-Bezout

factorization

((Do,N0) and

(~0,130)

respectively) and of matrices Yo,Zo,~o and Z0 such that:

-Zo

Yo 13o lq.o ] I-No ~o

oo ,o]=IoI 0,]

These state--space realizations are given by Do(s) = [DRA,DRB,DRC,DRD];

13o(S) -~ [DLA,DLB,DLC,DLD];

No(s)

[N~,NRB,NRC,NRD];

iqo(S) = [NLA,NLB,NLC,NLD];

Yo(s) = [YRA,YRB,YRC,YRD];

~o(S) --- [YLA,YLB,YLC,YLD];

Zo(s) = IZRA,ZRB,ZRC,ZRD];

~,o(S)

=

=

[ZLA,ZLB,Z1C,ZLD].

norcor is an abbreviated version of ncoprfac and produces only state-space realizations of the normalized right- and left-Bezout factorizations (Do,N0) and (lqo,I3o) of P.

103

Aleorithm : ncoprfac and norcor use the construction method for state--space realizations of normalized Bezout factorizations as given in theorem 4.1.1. Nested functions : sqrtl; see appendix E. See also : speccopr (see appendix B); this is a special purpose algorithm to be used in other algorithms. Pro gram~ function

Given a

:

[DRA,DRB33RC~DRD,NRA,NRB,NRC,NRD,.. YRA,YRB,YRC,YRD,ZRA,ZRB,ZRC,ZRD,.. DLA,DLB,DLC,DLD,NLA,NLB,NLC,NLD,.. YLA,YLB,YLC,YLD,7_.LA,7_Z,B,ZLC,ZLD]=ncopffac(A,B ,C,D) minimal

realization [A,B,C,DI of a transfermatrix P(s), ncoprfac computes realizations of a

normalized right coprime factorization

of P. P=ND-'I with D=[DRA,DRB,DRC,DRD1 and N=[NRA,NRB,

NRC,NRD]; Y=[YRA.YRB.YRC,YRD1 and Z=[ZRA,ZRB,ZRC,ZRD] are such that YD+ZN=I. The left cop~me case is treated in a similar way.

[m,nl=size(D); H=eye(n)+(D'*D);L-~ye(m)+(D*D'); Hinv=inv(l-I);Linv=inv(L); Ahulp 1=A--(B*Hinv*D'*C); Ahulp2=(A--(B*D'*Linv*C))'; [X,XPERR]--aresolv(Ahulp1,tY*Linv*C,B *I-Iinv*B','Schur'); [Y,YPERR]=aresolv(Ahulp2,B*Hinv*B',C*Linv*C,'Schur'); F=Hinv*(fD'*C)+fB'*X)); AC=A--(B*F); K=((B *D')+(Y*C'))*Linv; AO=A--(K*C); Hr=sqrtl(H);Hrinv--inv(Hr); Lr=-sqrtl (L) ;Lrinv--inv(Lr); D R A =A C ; D R B = - B *Hrinv,DR C=F;DRD=Hrinv;

NRA=AC;NRB=B *Hrinv;NRC=C-(D*F);NILD=D*Hrinv; Y R A =A O; Y R B = B - - ( K *D ); Y R C=Hr*F; YRD=I-~,

104

ZRA=AO;ZRB=K;ZRC---Hr*F;ZRD=zeros(n,m); DLA=AO;DLB--K;DLC---Lrinv*C;DLD--Lrinv; NLA=AO;NLB=B-(K*D);NLC=/.a'inv*C;NLD=Lrinv*D; YLA=AC;YLB=K*Lr,YLC=C-(D*F);YLD=Lr; ZLA=AC;ZLB=K*Lr,ZLC=F;ZLD=zeros(n,m);

function

IDRA,DRB,DRC,I)RD,NRA,NRB,NRC.NRD... DLA,DLB,DLC,DLD,NLA,NLB,NLC,NLD]=norcor(A,B,C,D)

Given a minimal realization [A,B.C,D] of a transfermatrix P(s), norcor computes reali2ations of a normalired right coprime factorization of P, (P--ND"1 with D=IDRA,DRB,DRC,DRD] and N=[NRA,NRB,NRC,NRD1) and a normalized

left

coprime

factorization

of

P

(P=D-IN

N=[NLA,NLB,NLC,NLD]).

[m,n]=size(D); H=eye(n)+(D'*D );L=eye(m)+(D*D');

Hinv=inv(H);Linv=inv(L); Alaulp1=A-(B*Hinv*D'*C); Ahulp2=(A---(B *D'*Linv*C))';

[X,XPERR]=aresolv(Ahulpl,C'*Linv*C,B*Hinv*B','Schur'); [Y,YPERR]=aresolv(Ahulp2,B*Hinv*B',C'*Linv*C,'Schur');' F=Hinv*((D'*C)+(B'*X)); AC=A-(B*F); K=((B*D')+(Y*C'))*Linv; AO=A-(K*C); Hr=sqrtl(H);Hrinv--inv(Hr); Lr=-sqrtl(L);Lrinv=inv(Lr); DRA=AC;DRB=-B*Hrinv;DRC=F;DRD=Hrinv; NRA=AC;NRB=B*Hrinv;NRC=C-(D*F);NRD=D*Hrinv; DLA=AO;DLB=K;DLC=-Lrinv*C;DLD=Lrinv; I~A=AO;NLB=B-(K*D);NLC=Lrinv*C;NLD=Lrinv*D;

with

D=[DLA,DLB,DLC,DLD]

and

105 spcccopr

Purpose : Normalized Bezout factorization of a plant P, and calculation of the matrix V, associated with P. speccopr is a special purpose algorithm, to be used in the functions robstab and robstgl. SvnoDsis : [DRA,DRB,DRC,DRD,NRA,NRB,NRC,NRD,.. YLA,YLB,YLC,YI_.D,ZLA,ZLB,ZLC,ZLD,.. VA,VB,VC,VD] = speccopr(A,B,C,D) Description : Given a minimal realization [A,B,C,D] of a transfermatrix P(s), speccopr computes state-space realizations of a normalized right-Bezout factorization (Do,No) of P and of matrices ~'0 and ~o such that (B.I) holds. (So speccopr delivers only a part of the output of ncoprfac.) Also a minimal state--space realization of the matrix V, associated with P as defined in (4.57), is given. These state-space realizations are returned as follows:

Do(s) = [DRA,DRB,DRC,DRD];

~0(s)

=

[YLA,YLB,YLC,YLD];

N0(s) -- [NRA,NRB,NRC,NRD];

~0(s)

=

[ZLA,ZLB,ZLC,ZLD];

V(s) = [VA,VB,VC,VD]. Algorithm : speccopr uses the construction method of theorem 4.1.I. for the state--space realizations of the matrices D0,N0,~ 0 and ~0, and the method of theorem 5.2.1. for the state--space realization of V (formula (5.4)). Nested functions : sqrtl; see appendix E. See also :

ncoprfac (see appendix B). Program : See

next page.

106

function

[DRA,DRB,DRC,DRD,NRA,NRB,NRC,NRD,.. YLA,YLB,YLC,YLD,ZI.,A,ZLB,ZLC,ZLD,.. VA,VB,VC,VD]=speccopr(A,B,C,D)

Given a mira'real realization [A,B,C,DI of a tramfermatrix P, there are an n.c.rf. (DdV) of P and an n.l.cf. (D,N) of P and stable matrices Y, Z and Y, Z, such that

f z ')I ~ ~] I' °1 speccopr

computes

Y=I'YLA,Yi.,B,YLC,YLD]

state and

space

realizations

~--[ZLA,TJ..B,ZLC,ZLD).

of

D=[DRA,ORB.DRC.DRD],

Moreover,

V = D*Z - N*Y (V=[VA,VB,VC,VDI) is computed.

[m,nl=sizefD);[p,ql=size(A); H=cye(n)+(D'*D);l.,=eye(m)+(D*D'); Hinv--inv(I-I);Linv=inv(L); Ahulp 1=A--
NRA=AC;NRB=B*Hrinv;NRC=C-(D*F);NRD=D*Hrinv; YLA=AC;YLB=K*Lr;YLC=C-(D*F);YLD=Lr; ZLA=AC;ZLB=K*Lr,ZLC=F;ZLD=zeros(n,m); VA=--AC';VB=(cyc(p)+(X*Y))*C'*Lrinv; VC=Hrinv*B';VD~--Hrinv*D'*Lr;

a

state

space

N=[NRA,NRB,NRC,NRD], realization

of

the

matrix

APPENDIX C This appendix contains func•n--descriptions of the MATLAB-functions hankel, hanneh and robstab.

hankel l:~ll~OSe

:

Computation of the norm of a Hankel---operator. Synopsis : norm = hankel(A,B,C,D) Description : Given a (not necessary minimal) realization [A,B,C,D] of a real--rational matrix V(s), hankel computes the norm of the Hankel--operator with symbol V. Algorithm : hankel uses the algorithm for the computation of the norm of a Hankel-operator as given in section 5.3. See als0 : hanneh (see appendix C); this function not only computes the norm of the Hankel--operator with symbol V, but also solves the related Nehari-problem. Proeram : function norm=hankel(A,B,C,D) Given

a

realization

[A,B ,C,D]

of

a

real

rational

matrix

V(s)

hankel-operatar with symbol V.

[A 1,B 1,C 1,D l,A4,B2,C2,D2,m]=stabproj(A,B,C,D); B BT2=B 2" (B2');CTC2=(C2')*C2; LC=Iyap(A4,A4',-BBT2);LO=lyap(A4',A4,-CTC2); LCLO=-LC*LO;eigenval=eig(LCLO);a=max(eigenval);norm=sqrt(a);

hankel

computes

the

norm

of

the

108

hanneh PUI'DOS¢ :

Computation of the norm of a Hankel-operator and of a sub-optimal solution to the related Nehari-problem. SYnopsis

:

[XA,XB,XC,XD,hannorm]

= hanneh(A,B,C,D,tol)

Descriotion : Given a rrfinimal realization [A,B,C,D] of a real--rational matrix V(s), hanneh computes: i)

hannorm; the norm of the Hankel---operator with symbol V.

ii)

a realization [XA,XB,XC,XD] of a matrix X(s) e M(IRI-I..) which is a sub-optimal solution to the related Nehari-problem, within the desired accuracy-level, given by 'tor. I.e. a matrix X(s) a M(~H®) is calculated such that:

II v - x l l < l l r

v II 0 + t o l )

second

stated

(cA)

Algorithm : hanneh

uses

the

algorithm

in

section

5.4.

for

the

solution

of

the

Nehari-problem. This algorithm is based upon the method for solving the Nehari--problem described in the same paragraph (this is the same method, Francis gives in [4, ch.5]). In this method

the

norm of the Hankel-operator is calculated automatically, actually it is an

intermediate result. See also : hannehgl; this function gives for an anti-stable matrix V(s)

an exact

solution to the

Nehari--problem (see appendix D). Proem

:

function[XA,XB,XC,XD,hannorm]=hanneh(A,B,C,D,tol) Given a minimal realization [A,B.C,D] of a real rational matrix V(s) hanneh computes: i)

hannarm: the norm of the hankel-operator with symbol V

ii)

a realization [XA.XB,XC.XD1 of a matrix X(s) such that the Hinf-norm of V-X smaller is than Imnnorra*(1 +tol )

[A1 ,B 1,C 1,D 1,A4,B2,C2,D2,m]=stabproj (A,B,C,D); BBT2=B2* (B2');CTC2=(C2')*C2;

109

LC--lyap(A4,A4',-BBT2); LO=lyap(A4',A4,--~rc2); LCLO=LC*LO;eigenval=eig(LCLO);a--max(eigenval); al=l/(a*(l+tol)*(l+tol)); [p,ql--size(LCLO); Ninv=eye(p)-(a l*LCLO);N=invfNinv); [r,s]=size(A);[k,l]=size(C); hulpI=al*N*LO*B2;hulp2=(-N')*BBT2; XA=[A zeros(r,(2*p));zeros(p,r) A4 hulp2;zeros(p,(r+p)) (-A4'--(hulpl*B2'))]; XB=[B;N'*B2;hulpl]; XC=[C - C 2 zeros(k,p)]; XD=D, hannorm=sqrt(a);

robstab Purpose : Computation of the maximal stability radius belonging to a plant P and a sub--optimally robust controller C. Synopsis : [CA,CB,CC,CD,botmd] = robstab(PA,PB,PC,PD,tol) Description : Given a minimal realization [PA.PB,PC,PD] of a plant P, robstab computes the maximal stability radius belonging to P (this value is stored in the variable bound). Also a state--space realization [CA,CB,CC,CD] of a sub--optimally robust controller C is calculated, which realizes the value of bound within the desired accuracy-level, given by tol. So not only P is stabilized by C, but also the systems in the ball around P, with radius (bound)~'(1-tol), i.e. the whole set

{ P~, • M(~(s)) ] 8(P,P~,) < (bound) * (i - tol) }

(C.2)

is stabilized by C. When the order of the original plant P is n, the order of the calculated compensator is 10n.

110

Algorithm : robstab is the computer--implementation of the algorithm for the solution of the problem of (sub)--optimally robust control as given in section 5.5. This means that the internal Nehari-problem is solved by the method of section 5.4., which is implemented in the function hanneh. See ~I~0 : robstgl, rbstglfar (see appendix D); these are other algorithms for the solution of the problem of optimally robust control. Nested functions : speccopr; see appendix B. hanneh; see appendix C. ffadd, tfinv, ffmult; see appendix E. Prq£ram : function[CA,CB ,CC,CD,bound]--robstab(PA,PB ,PC,PD,tol) Given plant P=[PAPB,PC.PDI, robstab calculates a compensator C=[CA,CB,CC,CDI that stabilizes P, and for each plant Q. such that d(P,Q) < bound*(1-tol), C is also a stabilizing compensator. The variable bound represents the maximal stability radius belonging to P, i.e. the radius of the largest ball around P that can be stabilized by one compensator. In this way C is a sub-optimal solution to the problem of optimally robust control. In the computation a Nehari-problem has to be solved, tol determines how accurate this problem has to be solved (tolerance in accuracy).

[DRA,DRB,DRC,DRD,NRA,NRB,NRC,NRD,YLA,YLB,.. YLC,YLD,ZLA,ZLB,ZLC,ZLD,VA,VB,VC,VD] =speccopr(PA,PB,PC,PD); [RA,RB,RC,RD,Ahan]=hanneh(VA,VB,VC,VD,tol); w g = s q n ( l + ( A h a n * A h a n ) ) ; bound= 1/wg; [I-I1,H2,H3,H4]--ffmult(DRA,DRB,DRC;DRD,RA,RB,-RC,-RD); [HS,H6,H7,H8] =ffadd (ZLA,ZLB,ZI.,C,ZLD,H 1,I-12,I-I3,H4); [Ill ,H2,H3,H4] =tfmult(NRA,NRB,NRC,NRD,RA,RB,RC,RD); [I-I9,H10,H 11 ,H 12]=tfadd(YLA,YLB,YLC,YLD,HI,H2,H3,H4); [l-I1,H2,H3,H4]=tfinv(H9M103-111,H 12); [CA,CB,CC,CD]=t fmult(H5,H6,H7,H8,H 1,H2,I-I3,I-I4);

APPHNDIX D This appendix contains function-descriptions of the MATLAB--functions hannehgl, robstgl and rbstglfar.

~.--ehgl Purpose : Computation of the norm of a Hankel-operator, and of an exact solution to the related Nehari-problem. Svnoosis :

[XA,XB,XC,XD,hannorm]

= hannehgl(A,B,C,D)

Descritnion : Given a minimal state-space realization [A,B,C,D] of a real-rational a n t i - s t a b l e matrix V(s), hannehgl computes: i)

hannorm, the norm of the Hankel-operator with symbol V.

ii)

a realization [XA,XB,XC,XD] of a matrix X(s) a M(~-I,,), which is an exact solution to the Nehari-problem for the malrix V(s). I.e. the matrix X ( s ) ¢ M(~r-l,,) has the property that

II v - x

(D.1)

II = II r v II.

When the order of the matrix V is n, the state-space realization of the matrix X is of the order n-1. Algorithm : hannehgl uses the method of Glover as described in section 6.2. to solve the Nehari---problem exactly. The realization of X is constructed without the use of balanced realizations, but with the method given in [14] and implemented in the MATLAB-function ohkapp.

This last

function is a standard function in the Robust Control Toolbox. See a|so : hanneh (see

appendix

Nehari--problem. Prokrram : See next page.

B);

this

algorithm

computes

a

sub-optimal

solution

to

the

112

function [XA,XB,XC,XD,harmorm]=hannehgl(A,B,C,D) Given a minimal realization [A,B,C,D] of a real anti-stable matrix V(s), hannehgl computes: i)

hannorm: the norm of the Hankel-operator with symbol V

ii)

a realization [XA,XB,XC,XD] of a stable matrix X(s). such that the Hinf-norm of V-X is equal to hannorm, and the order of X is smaller than the order of V

[AX,BX,CX,DX,AY,BY,CY,DY,aug] =ohkapp(-A,B,---C,D,1,0); XA=-AY;XB =BY;XC----CY;XD=DY; hannorm=aug(1,1);

robstgl Purp0~¢ : Computation of the maximal stability radius belonging to a plant P and an optimally robust controller C. Svnoosis : [CA,CB,CC,CD,bound] -- robstgl(PA,PB,PC,PD) Dcscdpti0n : Given a minimal realization [PA,PB,PC,PD] of a plant P, robstgl computes the maximal stability radius belonging to P (stored in the variable bound), and a state--space realization [CA,CB,CC,CD] of an optimally robust controller C that realizes this bound. This controller C not only stabilizes P, but also systems in the ball around P, with radius 'bound', i.e. the set { P2~ e Mt~(s)) [ 8(P,P~) < bound }.

(I).2)

When the order of the plant is n, the order of the calculated compensator C is 6n-2. Algorithm : robstgl uses the method of Glover for the solution of the problem of optimally robust control, as explained in section 6.2. So the internal Nehari-problem is exactly solved by the method of Glover, which is implemented in the function hannehgl. S¢¢ also : robstab (see appendix C), rbstglfar (see appendix D); these are other algorithms for the solution of the problem of optimally robust control.

113

Nested functions : speccopr; see appendix B. hannehgl; see appendix D. ffadd, tfinv, ffmult; see appendix E. ProlFram : function [CA,CB,CC,CD,bound] --robstgl(PA,PB ,PC,PD) Given plant P=[PA,PB,PC,PD], robstgl calculates a compensator C=[CA,CB,CC, CD1 that stabilizes P, and for each plant Q, such that d(P,Q)
C is also a stabilizing compensator. Moreover, C is chosen in such

a way that bound is maximal. Le. for each compensator C' that stabilizes P, the radius of the ball around P of planls also stabilized by C', is not greater than bound.

[DRA,DRB,DRC,DRD,NRA,NRB,NRC,NRD,YLA,YLB,.. YLC,YLD,ZLA,ZLB,ZLC,ZLD,VA,VB,VC,VD]=speccopr(PA,PB,PC,PD); [P.A,RB,RC,RD,Ahan] =hannehgI(VA,VB,VC,VD); wg=sqrt(1 +(Ahan*Ahan)); bound=l/wg; [H1,H2,H3.1-14]=tfmult(DRA,DRB,DRC,DRD,RA,RB,-RC,-RD); IH5,H6,H7,H8]=ffadd(ZLA,ZLB,ZLC,ZLD,H 1,I-I2,H3,H4); [H t,H2,H3,H4]=tfmult(NRA,NRB,NRC,NRD,RA,RB,RC,RD); [H9,H10,HI 1,H 12]=ffadd(YLA,YLB,YLC,YLD,H 1,H2,H3,H4); [H1,H2,H3,H4]=ffinv(H9,H 102-I11,H12); [CA,CB ,CC,CD] =t fmult(HS,H6,H7,H8,H1 ,H2,I-13,H4);

rbstglfar PUrDOS~ " Computation of the maximal stability radius belonging to a plant P and a sub--optimally robust controller C. Svnoosis

:

[CA,CB,CC,CD,bound] = rbstglfar(PA,PB,PC,PD,tol) Dcscrintion : Given a minimal realization [PA,PB,PC,PD] of a plant P, rbstglfar computes the maximal stability radius belonging to P (which is returned in the variable bound) and a state-space

114

realization [CA,CB,CC,CD] of a sub--optimally robust controller C, which realizes this bound within an accuracy-level of tol (tol is one of the input variables). This means that C stabilizes all the systems in the ball around P with radius (bound)*(1--tol), i.e. C stabilizes the set (I).3)

{ P~. • M(~(s)) I 8(P,P~) < (bound) * (l-tol) }. The order of this compensator C is equal to the order of the original plant P. A1~orithm

:

To compute the maximal stability radius belonging to P, rbstglfar uses the method of theorem 6.3.2. So the value of bound is calculated with formula (6.22). The state-space realization of C is constructed by the method of Glover and McFarlane as given in theorem 6.3.3. Formula (6.35) is used for the computation of the realization of C. See ialso : robstab (see appendix C), mbstgl (see appendix D); these are other algorithms for the solution of the problem of optimally robust control. FloEram

:

rune tion [CA,CB,CC, CD,bound] ----rbstglfar(PA,PB,PC,PD,tol) Given a minimal realization [PA,PB,PC.PD] of a plant P, rbstglfar calculates the maximal stability radius "bound" and

a

realization

[CA,CB,CC,CD1

of a

compensator C,

accuracy-level of toL

[m,n]=size(PD);[r,k]=size(PA); H=eye(n)+(PD'*PD);L=cye(m)+(PD*PD'); Hinv=inv(H);Linv=inv (L); Ahulpl =PA--~PB *Hinv *PD'*PC); Ahulp2=(PA-(PB *PD'*Linv*PC))'; [X,XPERR] =aresolv(Ahulp 1,PC'*Linv*PC,PB *Hinv*PB','Schur'); [Y,YPERR]=aresolv(Ahulp2,PB *Hinv*PB',Pe'*Linv*PC,'Schur'); l=max(eig(Y*X));bound=l/sqn(1 +l); gamma=(1/bound)*(l+tol);g2=gamma*gamma; F=Hinv*((PD'*PC)+(PB'*X));

which

realizes this bound within an

115

AC=PA-(PB*F); Wl=((l-g2)*eye(r))+(X*Y); CB=g2*(invfW1'))*Y'PC'; CA=AC+CB*(PC-(PD*F)); CC=PB'*X; CD=-PD';

APPENDIX E This appendix contains function--descriptions of the MATLAB-functions tfadd, tfinv, tfmult and sqrtl.

tfadd Pll/'DOSC :

Addition of two transfermatrices in state-space form. SVnODSiS

:

[P,Q,R,S] = ffadd(AI,BI,CI,DI,A2,B2,C2,D2) Description : Given state-space realizations of two transfermatrices

of the same size, tfadd computes a

state-space realization of the sum of these transfermatrices, i.e. [P,Q,R,S] = [A1,B1,C1,D1] + [A2,B2,C2,D2]. Algorithm : tfadd uses the state-space formula of page ix for the addition of two transfermatrices,

to

construct a state-space realization of the sum. P r o v~ r a m

:

function[P,Q,R,S]=tfadd(A I,B 1,C 1,D 1,A2,B2,C2,D2) Given

state

space

realizations

of two

transfermatrices

of the. same size,

realization of the sam of these transferma~riccs, i.e.

[P,QJt,SI=[A1,B1,C1,DlI+[A2,B2,C2,D21. [nal,mal]=size(A1); [na2,ma2]=size(A2); P=[A1 zeros(nal,ma2);zeros(na2,mal) A2]; Q=[BI;B2]; R=[CI C2]; S=DI+D2;

~'add calculates

a state

space

117 tfinv Puroose : Inversion of a transferrnatrix in state-space form. Svnovsis : [P,Q,R,S] = ffinv(A,B,C,D) Descrimion : Given

a

state-space

realization

of

a

square

invertible

tfinv

transfermatrix,

computes

a

transfermatrix,

to

state--space realization of the inverse of this transfermatrix, i.e, [P,Q,R,S] = [A,B,C,D]-I. Algorithm : tfinv uses

the

state--space

formula

of

page

ix

for

the

inversion

of

a

construct a state-space realization of the inverse. Program : function[P,Q,R,S]=tfinv(A,B,C,D) Given a state space realization of a square transfermatrix (with D invertible), realization of the inverse of tMs transfermatrix. So IP,Q,R,SI=inv([A,B.C,DI).

hulp=inv(D); P=A-(B*hulp*C); Q=B*hulp; R=---hulp*C; S=hulp;

~ n v calculates a state space

118

tfmult PtlrDose

:

Multiplication of two transfermatrices in state-space form. SYnoPsiS : [P,Q,R,S] = ffmuh(AI,BI,C1,DI,A2,B2,C2,D2) Descrintion : Given state-space realizations of two transfermatrices

of compatible size (the number of

columns of the first one equals the number of rows of the second one), tfmult ealculates a state-space realization of the product of these transfermatrices, i.e. [P,Q,R,S] = [A1,B1,C1,DI] * [A2,B2,C2,D2]. Al~orithm : tfmult uses the state-space formula of page ix for the multiplication of two transfermatrices, to construct a state-space realization of the product. P r o RT"dm :

function[P,Q,R,S]=tfmult(A1,B1,CI,D1,A2,B2,C2,D2) Given state space realizations of two transfermatrices (number of columns of the first equals the number of rows of the second), ~-rnult computes a state space realizoZion of the product of the two transfermatrices, Le. [P,Q.R,SI=IA1,B1,C1.D I I* [A2.B2,C2,D 21.

[nal,mal]=size(A1); [na2,ma2]=size(A2); P=[A1 (Bl*C2.);zeros(na2,mal) A2]; Q=[(BI*D2);B2]; R=[CI (DI*C2)]; S=DI*D2;

119 sqrt] Puroose : Computation of the square root of a positive definite matrix. Svnousis : root = sqrtl(A) Descdotion : Given a positive definite matrix A, sqrtl computes the matrix A t12 with help of the singular value decomposition. So when A is given by A--USV

T,

with S diagonal, then sqrtl(A) is sqnl(A) = U S I/2 V T. Program : function wortcl=sqnI(A) sqnl computes the square root of a PD matrix

[U,S,V]=svd(A); D=S.^(1/2); wortel=U*D*V';

with help of the SV-Decomposition.

APPENDIX F This appendix contains solutions to the problem of (sub)--optimally robust stabilization for a 1

pIant with transferfunction P(s) = s--:l" ' calculated in five different ways:

i)

with the MATLAB-function robstab; tolerance level tol = 10-5;

ii)

with the MATLAB-funetion robstab; tolerance level tol = 10-8;

iii)

with the MATLAB-funetion robstgl;

iv)

with the MATLAB-function rbstglfa~, tolerance level tol = 10-5;

v)

with the MATLAB-function rbstglfar;, tolerance level tol = 10-~.

In each case the calculated maximal stability radius is given and a state--space realization of a compensator that realizes this bound within (for the eases i), ii), iv) and v) ) the desired accuracy level.

i) MATLAB-function robstab, tolerance level tol = 10-5 PA=

1

PB = I PC=

I

PD=0

bound = 0.3827

tol =

t.0000¢-005

121

CA =

1.0e+006 *

Columns 1 through 7

-0.0000 0 0 0 0 0 0 0 0 0

0 -0.00~ 0 0 0 0 0 0 0 0

0 0.0000 0.0000 0 0 0 0 0 0 0

0 -0.0000 0 0.0000 0 0 0 0 0 0

0 0 0 -2.3314 --0.1414 0 0 0 0 0

-0.0000 0 -0.0000 -0.3414 -0.0207 -0.0000 0 -0.0000 -0.3414 -0.0207

-0.0000 0 -0.0000 -0.3414 --0.0207 -0.0000 -0.0000 -0.0000 -0.3414 -0.0207

1.0000

0

0

0

Columns 8 through 10 0 0 0 0 0 0 0.0000 0.0000 0 0 CB =

0 0 0 0 0 0 0 0 0 0 0 0 -0.0000 0 0 0 0.000(3 -2.3314 0 -0.1414

1.0e+005 * 0.0000 0 0.0001 3.4143 0.2071 0.0000 0 0.0001 3.4143 0.2071

CC = Columns 1 through 7 2.4142

2.4142

--1.0000

Columns 8 through 10 0 CD=0

0

0

122 ii) MA'II.,AB---fimcfion robstab, tolerance level tol = 10-s PA= 1

PB=I PC=I PD=0 bound = 0.3827 tol = 1.000(O--008 CA =

1.0e+009 *

Columns 1 through 7 -0.0000 0 0 0 0 0 0 0 0 0

0 -0.0000 0 0 0 0 0 0 0 0

0 0.0000 0.0000 0 0 0 0 0 0 0

Columns 8 through 10 0 0 0 0 0 0 0.0000 0.00~ 0 0 CB =

0 0 0 0 0 0 -0.0000 0 0.0000 0

1.0e+008 *

0.~ 0 0.0000 3.4142 0.207t 0.0000 0 0.0000 3.4142 0.2071

0 0 0 0 0 0 0 0 --2.3314 -0.1414

0 -0.0000 0 0.0000 0 0 0 0 0 0

0 0 0 -2.3314 -0.1414 0 0 0 0 0

-0.00130 0 -0.0000 -0.3414 -0.0207 -0.0000 0 -0.OOOO -0.3414 -0.0207

-0.0000 0 -0.0000 -0.3414 -0.0207 -0.0000 -0.0000 -0.0000 --0.3414 -0.0207

123 CC = Columns 1 through 7 2.4142

2.4142

-1.0000

1.0000

Columns 8 through 10 0

0

0

CI).~ 0

iii) MATLAB----function robstgl PA = 1

PB= 1 PC= 1 PD=O bound = 0.3827 CA = --1.4142 0 0 0

0 -1.4142 0 0

--2.4142 2.4142 --3.8284 2.4142

-2.4142 2.4142 -2.4142 1.0000

2.4142

-2.4142

--2.4142

CB = 2.4142 -2.4142 2.4142 -2.4142 CC= 2.4142

CD=

2,4142

124 iv) M

PA= PB

A

~

o

n

rbstglfar, tolerance level tol = 1 0 s

1

= 1

PC=I PD=0

b o u n d -- 0.3827

tol = 1.O000e-O05

CA = -1.2071e+005 CB = -1.2071e+005 C C = 2.4142 CD=0

v) M A T L A B - f u n c f i o n rbstglfar, tolerance level tol = 10 ~

PA= 1 PB = 1 PC=I PD=0

b o u n d = 0.3827

tol

= 1.0000e-008

C A = --1.2071 e + 0 0 8 CB = -1.2071e+008 C C = 2.4142 CD=0

REFERENCES [1] [2] [3] [4] [5] [6] [7] [8]

[9] [10] [11] [12]

[13] [14]

[15]

N.A. Bruinsma and M. Steinbuch, A fast algorithm to compute the H®-norm of a transferfunction. Philips Research Laboratories Eindhoven. Eindhoven, 1989. R. Chiang and M. Safonov, Robust-Control Toolbox, User's Guide. The MathWorks, Inc. South Natick, 1988. H.O. Cordes and J.P. Labrousse, The invariance of the index in the metric space of closed operators. Journal of Mathematics and Mechanics, vol. !2, pp. 693-719, 1963. B.A. Francis, A Course in H Control Theory. Berlin-Heidelberg-New York, Springer Veflag, 1987. Lecture Notes in Control and Information Sciences, vol. 88. T.T. Georgiou, On the computation of the gap metric, Systems and Control Letters, vol. 11, pp. 253-257, 1988. T.T. Georgiou and M.C. Smith, Optimal Robustness in the Gap Metric. Preprint, 1989. K. Glover, All optimal Hankel-norm approximations of linear multivariable systems and their L**--error bounds. Int. J. Control, vol. 39, pp. 1115-1193, 1984. K. Glover and D. McFarIane, Robust stabilization of normalized eopfime factor plant descriptions with H**--bounded uncertainty. IEEE Trans. on Automat. Control, vol. 34, pp. 821-830, 1989. T. Kato, Perturbation Theory for Linear Operators. Berlin-New York, Springer Verlag, 1966. M.A. Krasnosel'skii, G.M. Vainikko and P.P. Zabreiko, Approximate Solution to Operator Equations. Groningen, Wolters-Noordhoff, 1972. E. Kreyszig, Introductory Functional Analysis with Applications. New York, Wiley, 1978. D. Meyer and G. Franklin, A connection between normalized coprime factofizations and linear quadratic regulator theory. IEEE Trans. on Automat. Control, vol. AC-32, pp. 227-228, 1987. Z. Nehari, On bounded bilinear forms. Annals of Mathematics, vol. 65, pp. 153-162, 1957. M. Safonov, R. Chiang and D. Limebeer, Hankel model reduction without balancing A descriptor approach. Proc. of the 26 a~ Conf. on Decision and Control, Los Angeles, 1987.

M. Vidyasagar, Control System Synthesis: A Factorization Approach. Cambridge, MA, M.I.T. Press, 1985. [16] M. Vidyasagar, Normalized coprime factorizations for nonstrictly proper systems. IEEE Trans. on Automat. Control, vol. 33, pp. 300---301, 1988. [17] M. Vidyasagar and H. Kimura, Robust controllers for uncertain linear multivariable systems. Automatica, vol. 22, pp. 85-94, 1986. [18] S.Q. Zhu, Graph topology and gap topology for unstable plants. IEEE Trans. on Automat. Control, vol. 34, pp. 848-855, 1989.

126

[i9]

S . Q . Zhu, Robustness of Feedback Stabilization: A

[20]

dissertation, Eindhoven University of Technology, Eindhoven, 1989. S.Q. Zhu, M.L.J. Hautus and C. Praagman, Sufficient conditions for robust BIBO stabilization: Given by the gap metric. Systems and Control Letters, vol. 11, pp. 53-59, 1988.

topological approach. Ph. D.